diff --git a/.gitattributes b/.gitattributes index e39e98722095c8f88f34ef8b8999c934954fe3e2..a26569e55df358e22e72ec271791611ef96985a1 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5098,3 +5098,13 @@ lambda0.001/elic-featurecoding-8bit-individual/fc_hellaswag/sample119-layer4-ite lambda0.001/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample56-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text lambda0.001/hyperprior-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample122-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text lambda0.001/hyperprior-featurecoding-8bit-individual/fc_winogrande/sample134-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_arc_challenge/sample49-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_arc_challenge/sample73-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_arc_challenge/sample92-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_arc_challenge/sample96-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample1087-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample2-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample543-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample778-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample91-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample52-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/dtufc_hyperprior-featurecoding_qwen_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/dtufc_hyperprior-featurecoding_qwen_individual.log index 33288e47b4358255b399b7adc5bfbb021ff22607..14aa3db994ec964eb42c13295eb68262133a3641 100644 --- a/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/dtufc_hyperprior-featurecoding_qwen_individual.log +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/dtufc_hyperprior-featurecoding_qwen_individual.log @@ -1,16958 +1,3 @@ -Experiment: dtufc_hyperprior-featurecoding_qwen_individual -Log file: output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/dtufc_hyperprior-featurecoding_qwen_individual.log -DTUFCCodecConfig: - arch: hyperprior-featurecoding - handler: qwen - checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar - transform_type: kmeans - transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json - bit_depth: 8 - device: cuda:0 -Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar -Checkpoint epoch: 559 -Loaded hyperprior-featurecoding (1-channel) on cuda:0 -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json -Loaded per-key mappings: model=qwen - Keys: ['layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache', 'layer.4.output'] ----------------- ------------------------------------------------------------------------------------------------------------------------------ -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding -Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar -Transform type kmeans -Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json -Input ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k -Output output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k ----------------- ------------------------------------------------------------------------------------------------------------------------------ -Files found: 100 ----------------------------------------------------------------------- - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1-layer4-item1.zst (1/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 173, 128) -Output shape: (1, 173, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.output: torch.Size([1, 173, 4096]) -> torch.Size([1, 1, 173, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,848B, BPFP=0.2189 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 31,352B, BPFP=1.4158 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 16,112B, BPFP=0.7276 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 32,440B, BPFP=1.4650 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,336B, BPFP=0.8280 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 32,876B, BPFP=1.4846 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,656B, BPFP=0.8876 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 31,792B, BPFP=1.4357 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 15,160B, BPFP=0.6846 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 33,280B, BPFP=1.5029 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 82,504B, BPFP=0.9314 -⌛️ [2/4] FRONTEND: Frontend time: 0.632s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.432s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02440321 19.99438900 - layer.0.v_cache 0.00000027 0.00030328 - layer.1.k_cache 0.00314874 1.40342324 - layer.1.v_cache 0.00000072 0.00104232 - layer.2.k_cache 0.00114712 0.71405091 - layer.2.v_cache 0.00000110 0.00156651 - layer.3.k_cache 0.00139354 0.80315139 - layer.3.v_cache 0.00000204 0.00260795 - layer.4.k_cache 0.00353492 1.63246075 - layer.4.v_cache 0.00000301 0.00429316 - layer.4.output 0.00018661 0.10863270 - ------------------------------------------------------------------------------------- - TOTAL 0.00245579 1.78512995 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 318356 -BPFP 1.0269 bits/point -EBPFP 2.0538 equivalent bits/point -MSE 1.785130 ----------------------- -------------------------------------------------------- -Time: 1.075s Load: 0.011s, Pack+Encode: 0.632s, Decode+Unpack: 0.432s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7851 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample1-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample100-layer4-item1.zst (2/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample100-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 109, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 109, 128) -Output shape: (1, 109, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.0.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.1.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.1.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.2.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.2.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.3.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.3.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.4.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.4.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.4.output: torch.Size([1, 109, 4096]) -> torch.Size([1, 1, 109, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,180B, BPFP=0.2279 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,284B, BPFP=1.5972 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,272B, BPFP=0.8079 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,160B, BPFP=1.5883 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,532B, BPFP=0.9699 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,620B, BPFP=1.6213 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,972B, BPFP=1.0014 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,616B, BPFP=1.5493 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,552B, BPFP=0.8280 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,660B, BPFP=1.6241 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 55,304B, BPFP=0.9910 -⌛️ [2/4] FRONTEND: Frontend time: 15.091s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 109, 128]) - layer.0.v_cache: torch.Size([1, 8, 109, 128]) - layer.1.k_cache: torch.Size([1, 8, 109, 128]) - layer.1.v_cache: torch.Size([1, 8, 109, 128]) - layer.2.k_cache: torch.Size([1, 8, 109, 128]) - layer.2.v_cache: torch.Size([1, 8, 109, 128]) - layer.3.k_cache: torch.Size([1, 8, 109, 128]) - layer.3.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.k_cache: torch.Size([1, 8, 109, 128]) - layer.4.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.output: torch.Size([1, 109, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.305s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 109, 128]) - layer.0.v_cache: torch.Size([1, 8, 109, 128]) - layer.1.k_cache: torch.Size([1, 8, 109, 128]) - layer.1.v_cache: torch.Size([1, 8, 109, 128]) - layer.2.k_cache: torch.Size([1, 8, 109, 128]) - layer.2.v_cache: torch.Size([1, 8, 109, 128]) - layer.3.k_cache: torch.Size([1, 8, 109, 128]) - layer.3.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.k_cache: torch.Size([1, 8, 109, 128]) - layer.4.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.output: torch.Size([1, 109, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02465495 23.46488407 - layer.0.v_cache 0.00000027 0.00031249 - layer.1.k_cache 0.00330346 1.59267992 - layer.1.v_cache 0.00000080 0.00107929 - layer.2.k_cache 0.00115767 0.74631325 - layer.2.v_cache 0.00000114 0.00155742 - layer.3.k_cache 0.00132842 0.84113676 - layer.3.v_cache 0.00000211 0.00254402 - layer.4.k_cache 0.00335301 1.70445153 - layer.4.v_cache 0.00000290 0.00414654 - layer.4.output 0.00024947 0.11996464 - ------------------------------------------------------------------------------------- - TOTAL 0.00248590 2.05992599 - (elements=1,562,624) ----------------------- --------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- --------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- --------------------------------------------------------- -Total Elements 1562624 -Total Bytes 220152 -BPFP 1.1271 bits/point -EBPFP 2.2542 equivalent bits/point -MSE 2.059926 ----------------------- --------------------------------------------------------- -Time: 15.404s Load: 0.008s, Pack+Encode: 15.091s, Decode+Unpack: 0.305s ----------------------- --------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 109, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0599 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample100-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample100-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample108-layer4-item1.zst (3/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample108-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,724B, BPFP=0.2194 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,328B, BPFP=1.5567 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,152B, BPFP=0.8177 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,724B, BPFP=1.6691 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,944B, BPFP=1.0425 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,460B, BPFP=1.6479 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,272B, BPFP=1.0689 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,844B, BPFP=1.5983 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,068B, BPFP=0.8109 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,352B, BPFP=1.7197 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 50,756B, BPFP=1.0220 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02579871 24.96814594 - layer.0.v_cache 0.00000028 0.00032650 - layer.1.k_cache 0.00331373 1.70889597 - layer.1.v_cache 0.00000086 0.00119817 - layer.2.k_cache 0.00113316 0.77987152 - layer.2.v_cache 0.00000115 0.00171556 - layer.3.k_cache 0.00137718 0.85014642 - layer.3.v_cache 0.00000224 0.00280789 - layer.4.k_cache 0.00334605 1.71824709 - layer.4.v_cache 0.00000324 0.00478564 - layer.4.output 0.00031880 0.13371406 - ------------------------------------------------------------------------------------- - TOTAL 0.00258942 2.18364264 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 201624 -BPFP 1.1599 bits/point -EBPFP 2.3199 equivalent bits/point -MSE 2.183643 ----------------------- -------------------------------------------------------- -Time: 0.504s Load: 0.008s, Pack+Encode: 0.208s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1836 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample108-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample108-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1087-layer4-item1.zst (4/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1087-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 50, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 50, 128) -Output shape: (1, 50, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.0.v_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.1.k_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.1.v_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.2.k_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.2.v_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.3.k_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.3.v_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.4.k_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.4.v_cache: torch.Size([1, 8, 50, 128]) -> torch.Size([1, 1, 50, 1024]) - layer.4.output: torch.Size([1, 50, 4096]) -> torch.Size([1, 1, 50, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,832B, BPFP=0.2863 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 8,380B, BPFP=1.3094 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 4,584B, BPFP=0.7163 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 9,316B, BPFP=1.4556 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 5,640B, BPFP=0.8812 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 9,572B, BPFP=1.4956 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 5,780B, BPFP=0.9031 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 9,080B, BPFP=1.4187 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 4,128B, BPFP=0.6450 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 9,480B, BPFP=1.4812 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 19,272B, BPFP=0.7528 -⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 50, 128]) - layer.0.v_cache: torch.Size([1, 8, 50, 128]) - layer.1.k_cache: torch.Size([1, 8, 50, 128]) - layer.1.v_cache: torch.Size([1, 8, 50, 128]) - layer.2.k_cache: torch.Size([1, 8, 50, 128]) - layer.2.v_cache: torch.Size([1, 8, 50, 128]) - layer.3.k_cache: torch.Size([1, 8, 50, 128]) - layer.3.v_cache: torch.Size([1, 8, 50, 128]) - layer.4.k_cache: torch.Size([1, 8, 50, 128]) - layer.4.v_cache: torch.Size([1, 8, 50, 128]) - layer.4.output: torch.Size([1, 50, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.208s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 50, 128]) - layer.0.v_cache: torch.Size([1, 8, 50, 128]) - layer.1.k_cache: torch.Size([1, 8, 50, 128]) - layer.1.v_cache: torch.Size([1, 8, 50, 128]) - layer.2.k_cache: torch.Size([1, 8, 50, 128]) - layer.2.v_cache: torch.Size([1, 8, 50, 128]) - layer.3.k_cache: torch.Size([1, 8, 50, 128]) - layer.3.v_cache: torch.Size([1, 8, 50, 128]) - layer.4.k_cache: torch.Size([1, 8, 50, 128]) - layer.4.v_cache: torch.Size([1, 8, 50, 128]) - layer.4.output: torch.Size([1, 50, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02911560 36.89193359 - layer.0.v_cache 0.00000029 0.00034396 - layer.1.k_cache 0.00403332 2.31847778 - layer.1.v_cache 0.00000087 0.00121320 - layer.2.k_cache 0.00117108 0.89625557 - layer.2.v_cache 0.00000108 0.00168151 - layer.3.k_cache 0.00141424 1.00498077 - layer.3.v_cache 0.00000205 0.00276785 - layer.4.k_cache 0.00319313 2.10868530 - layer.4.v_cache 0.00000283 0.00427966 - layer.4.output 0.00027592 0.19373522 - ------------------------------------------------------------------------------------- - TOTAL 0.00285987 3.14325429 - (elements=716,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 716800 -Total Bytes 87064 -BPFP 0.9717 bits/point -EBPFP 1.9434 equivalent bits/point -MSE 3.143254 ----------------------- -------------------------------------------------------- -Time: 0.433s Load: 0.005s, Pack+Encode: 0.220s, Decode+Unpack: 0.208s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 50, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 50, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 3.1433 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1087-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample1087-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1128-layer4-item1.zst (5/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1128-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 57, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.004s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 57, 128) -Output shape: (1, 57, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.0.v_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.1.k_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.1.v_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.2.k_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.2.v_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.3.k_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.3.v_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.4.k_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.4.v_cache: torch.Size([1, 8, 57, 128]) -> torch.Size([1, 1, 57, 1024]) - layer.4.output: torch.Size([1, 57, 4096]) -> torch.Size([1, 1, 57, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,032B, BPFP=0.2785 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 9,376B, BPFP=1.2851 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 5,288B, BPFP=0.7248 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 10,252B, BPFP=1.4052 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 6,236B, BPFP=0.8547 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 10,664B, BPFP=1.4616 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 6,552B, BPFP=0.8980 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 10,424B, BPFP=1.4287 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 4,832B, BPFP=0.6623 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 10,548B, BPFP=1.4457 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 22,304B, BPFP=0.7643 -⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 57, 128]) - layer.0.v_cache: torch.Size([1, 8, 57, 128]) - layer.1.k_cache: torch.Size([1, 8, 57, 128]) - layer.1.v_cache: torch.Size([1, 8, 57, 128]) - layer.2.k_cache: torch.Size([1, 8, 57, 128]) - layer.2.v_cache: torch.Size([1, 8, 57, 128]) - layer.3.k_cache: torch.Size([1, 8, 57, 128]) - layer.3.v_cache: torch.Size([1, 8, 57, 128]) - layer.4.k_cache: torch.Size([1, 8, 57, 128]) - layer.4.v_cache: torch.Size([1, 8, 57, 128]) - layer.4.output: torch.Size([1, 57, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.204s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 57, 128]) - layer.0.v_cache: torch.Size([1, 8, 57, 128]) - layer.1.k_cache: torch.Size([1, 8, 57, 128]) - layer.1.v_cache: torch.Size([1, 8, 57, 128]) - layer.2.k_cache: torch.Size([1, 8, 57, 128]) - layer.2.v_cache: torch.Size([1, 8, 57, 128]) - layer.3.k_cache: torch.Size([1, 8, 57, 128]) - layer.3.v_cache: torch.Size([1, 8, 57, 128]) - layer.4.k_cache: torch.Size([1, 8, 57, 128]) - layer.4.v_cache: torch.Size([1, 8, 57, 128]) - layer.4.output: torch.Size([1, 57, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02901734 31.88975980 - layer.0.v_cache 0.00000028 0.00030999 - layer.1.k_cache 0.00371865 1.78879025 - layer.1.v_cache 0.00000081 0.00110517 - layer.2.k_cache 0.00114644 0.76529024 - layer.2.v_cache 0.00000109 0.00155992 - layer.3.k_cache 0.00141512 0.86198720 - layer.3.v_cache 0.00000208 0.00253129 - layer.4.k_cache 0.00325023 1.82629743 - layer.4.v_cache 0.00000285 0.00406325 - layer.4.output 0.00023785 0.15066522 - ------------------------------------------------------------------------------------- - TOTAL 0.00282188 2.69602539 - (elements=817,152) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 817152 -Total Bytes 98508 -BPFP 0.9644 bits/point -EBPFP 1.9288 equivalent bits/point -MSE 2.696025 ----------------------- -------------------------------------------------------- -Time: 0.364s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.204s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 57, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 57, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.6960 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1128-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample1128-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample117-layer4-item1.zst (6/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample117-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 113, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 113, 128) -Output shape: (1, 113, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.0.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.1.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.1.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.2.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.2.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.3.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.3.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.4.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.4.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.4.output: torch.Size([1, 113, 4096]) -> torch.Size([1, 1, 113, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,260B, BPFP=0.2254 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,532B, BPFP=1.4887 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,296B, BPFP=0.7118 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,420B, BPFP=1.5501 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,084B, BPFP=0.9046 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 23,148B, BPFP=1.6004 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,628B, BPFP=0.9422 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,180B, BPFP=1.5335 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,012B, BPFP=0.7613 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,540B, BPFP=1.6275 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 65,272B, BPFP=1.1282 -⌛️ [2/4] FRONTEND: Frontend time: 0.212s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 113, 128]) - layer.0.v_cache: torch.Size([1, 8, 113, 128]) - layer.1.k_cache: torch.Size([1, 8, 113, 128]) - layer.1.v_cache: torch.Size([1, 8, 113, 128]) - layer.2.k_cache: torch.Size([1, 8, 113, 128]) - layer.2.v_cache: torch.Size([1, 8, 113, 128]) - layer.3.k_cache: torch.Size([1, 8, 113, 128]) - layer.3.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.k_cache: torch.Size([1, 8, 113, 128]) - layer.4.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.output: torch.Size([1, 113, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.290s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 113, 128]) - layer.0.v_cache: torch.Size([1, 8, 113, 128]) - layer.1.k_cache: torch.Size([1, 8, 113, 128]) - layer.1.v_cache: torch.Size([1, 8, 113, 128]) - layer.2.k_cache: torch.Size([1, 8, 113, 128]) - layer.2.v_cache: torch.Size([1, 8, 113, 128]) - layer.3.k_cache: torch.Size([1, 8, 113, 128]) - layer.3.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.k_cache: torch.Size([1, 8, 113, 128]) - layer.4.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.output: torch.Size([1, 113, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03016301 25.10514035 - layer.0.v_cache 0.00000028 0.00032311 - layer.1.k_cache 0.00339028 1.58558182 - layer.1.v_cache 0.00000081 0.00111045 - layer.2.k_cache 0.00113693 0.73706318 - layer.2.v_cache 0.00000114 0.00159520 - layer.3.k_cache 0.00135151 0.81715251 - layer.3.v_cache 0.00000219 0.00266700 - layer.4.k_cache 0.00328814 1.61582123 - layer.4.v_cache 0.00000314 0.00448204 - layer.4.output 0.00017637 0.11754937 - ------------------------------------------------------------------------------------- - TOTAL 0.00286021 2.16722389 - (elements=1,619,968) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1619968 -Total Bytes 229372 -BPFP 1.1327 bits/point -EBPFP 2.2654 equivalent bits/point -MSE 2.167224 ----------------------- -------------------------------------------------------- -Time: 0.511s Load: 0.008s, Pack+Encode: 0.212s, Decode+Unpack: 0.290s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 113, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1672 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample117-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample117-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample120-layer4-item1.zst (7/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample120-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 108, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 108, 128) -Output shape: (1, 108, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.0.v_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.1.k_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.1.v_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.2.k_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.2.v_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.3.k_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.3.v_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.4.k_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.4.v_cache: torch.Size([1, 8, 108, 128]) -> torch.Size([1, 1, 108, 1024]) - layer.4.output: torch.Size([1, 108, 4096]) -> torch.Size([1, 1, 108, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,084B, BPFP=0.2231 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,052B, BPFP=1.5229 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,484B, BPFP=0.7584 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 21,896B, BPFP=1.5839 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,128B, BPFP=0.9497 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,108B, BPFP=1.5992 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,912B, BPFP=1.0064 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,492B, BPFP=1.5547 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,836B, BPFP=0.7839 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,516B, BPFP=1.6288 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 49,132B, BPFP=0.8885 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 108, 128]) - layer.0.v_cache: torch.Size([1, 8, 108, 128]) - layer.1.k_cache: torch.Size([1, 8, 108, 128]) - layer.1.v_cache: torch.Size([1, 8, 108, 128]) - layer.2.k_cache: torch.Size([1, 8, 108, 128]) - layer.2.v_cache: torch.Size([1, 8, 108, 128]) - layer.3.k_cache: torch.Size([1, 8, 108, 128]) - layer.3.v_cache: torch.Size([1, 8, 108, 128]) - layer.4.k_cache: torch.Size([1, 8, 108, 128]) - layer.4.v_cache: torch.Size([1, 8, 108, 128]) - layer.4.output: torch.Size([1, 108, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.287s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 108, 128]) - layer.0.v_cache: torch.Size([1, 8, 108, 128]) - layer.1.k_cache: torch.Size([1, 8, 108, 128]) - layer.1.v_cache: torch.Size([1, 8, 108, 128]) - layer.2.k_cache: torch.Size([1, 8, 108, 128]) - layer.2.v_cache: torch.Size([1, 8, 108, 128]) - layer.3.k_cache: torch.Size([1, 8, 108, 128]) - layer.3.v_cache: torch.Size([1, 8, 108, 128]) - layer.4.k_cache: torch.Size([1, 8, 108, 128]) - layer.4.v_cache: torch.Size([1, 8, 108, 128]) - layer.4.output: torch.Size([1, 108, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02569981 25.47918701 - layer.0.v_cache 0.00000028 0.00030983 - layer.1.k_cache 0.00351470 1.60561187 - layer.1.v_cache 0.00000073 0.00105341 - layer.2.k_cache 0.00116483 0.74965272 - layer.2.v_cache 0.00000104 0.00152671 - layer.3.k_cache 0.00137767 0.83545840 - layer.3.v_cache 0.00000202 0.00257442 - layer.4.k_cache 0.00329212 1.73496063 - layer.4.v_cache 0.00000308 0.00422881 - layer.4.output 0.00021651 0.13031950 - ------------------------------------------------------------------------------------- - TOTAL 0.00256588 2.20970299 - (elements=1,548,288) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1548288 -Total Bytes 209640 -BPFP 1.0832 bits/point -EBPFP 2.1664 equivalent bits/point -MSE 2.209703 ----------------------- -------------------------------------------------------- -Time: 0.500s Load: 0.007s, Pack+Encode: 0.206s, Decode+Unpack: 0.287s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 108, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 108, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2097 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample120-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample120-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1295-layer4-item1.zst (8/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1295-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 47, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.004s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 47, 128) -Output shape: (1, 47, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.0.v_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.1.k_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.1.v_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.2.k_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.2.v_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.3.k_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.3.v_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.4.k_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.4.v_cache: torch.Size([1, 8, 47, 128]) -> torch.Size([1, 1, 47, 1024]) - layer.4.output: torch.Size([1, 47, 4096]) -> torch.Size([1, 1, 47, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,820B, BPFP=0.3025 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 8,108B, BPFP=1.3477 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 4,492B, BPFP=0.7467 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 8,940B, BPFP=1.4860 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 5,360B, BPFP=0.8910 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 9,152B, BPFP=1.5213 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 5,732B, BPFP=0.9528 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 8,800B, BPFP=1.4628 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 4,048B, BPFP=0.6729 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 9,068B, BPFP=1.5073 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 18,372B, BPFP=0.7635 -⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 47, 128]) - layer.0.v_cache: torch.Size([1, 8, 47, 128]) - layer.1.k_cache: torch.Size([1, 8, 47, 128]) - layer.1.v_cache: torch.Size([1, 8, 47, 128]) - layer.2.k_cache: torch.Size([1, 8, 47, 128]) - layer.2.v_cache: torch.Size([1, 8, 47, 128]) - layer.3.k_cache: torch.Size([1, 8, 47, 128]) - layer.3.v_cache: torch.Size([1, 8, 47, 128]) - layer.4.k_cache: torch.Size([1, 8, 47, 128]) - layer.4.v_cache: torch.Size([1, 8, 47, 128]) - layer.4.output: torch.Size([1, 47, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.202s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 47, 128]) - layer.0.v_cache: torch.Size([1, 8, 47, 128]) - layer.1.k_cache: torch.Size([1, 8, 47, 128]) - layer.1.v_cache: torch.Size([1, 8, 47, 128]) - layer.2.k_cache: torch.Size([1, 8, 47, 128]) - layer.2.v_cache: torch.Size([1, 8, 47, 128]) - layer.3.k_cache: torch.Size([1, 8, 47, 128]) - layer.3.v_cache: torch.Size([1, 8, 47, 128]) - layer.4.k_cache: torch.Size([1, 8, 47, 128]) - layer.4.v_cache: torch.Size([1, 8, 47, 128]) - layer.4.output: torch.Size([1, 47, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02987218 36.43965052 - layer.0.v_cache 0.00000031 0.00035754 - layer.1.k_cache 0.00397950 2.17792105 - layer.1.v_cache 0.00000076 0.00116088 - layer.2.k_cache 0.00125400 0.84629018 - layer.2.v_cache 0.00000109 0.00167593 - layer.3.k_cache 0.00146534 0.97327715 - layer.3.v_cache 0.00000206 0.00282937 - layer.4.k_cache 0.00325167 2.14651976 - layer.4.v_cache 0.00000289 0.00439753 - layer.4.output 0.00022000 0.16748903 - ------------------------------------------------------------------------------------- - TOTAL 0.00290784 3.09028829 - (elements=673,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 673792 -Total Bytes 83892 -BPFP 0.9961 bits/point -EBPFP 1.9921 equivalent bits/point -MSE 3.090288 ----------------------- -------------------------------------------------------- -Time: 0.363s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 47, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 47, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 3.0903 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample1295-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample1295-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample13-layer4-item1.zst (9/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample13-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 154, 128) -Output shape: (1, 154, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.output: torch.Size([1, 154, 4096]) -> torch.Size([1, 1, 154, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,560B, BPFP=0.2313 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 29,240B, BPFP=1.4834 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,928B, BPFP=0.7573 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 31,340B, BPFP=1.5899 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,896B, BPFP=0.9586 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 32,080B, BPFP=1.6274 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 20,420B, BPFP=1.0359 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 32,148B, BPFP=1.6309 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 15,172B, BPFP=0.7697 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 32,976B, BPFP=1.6729 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 85,224B, BPFP=1.0809 -⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.385s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02515895 22.24257591 - layer.0.v_cache 0.00000027 0.00029764 - layer.1.k_cache 0.00328501 1.57980396 - layer.1.v_cache 0.00000091 0.00108004 - layer.2.k_cache 0.00115449 0.72744801 - layer.2.v_cache 0.00000127 0.00158661 - layer.3.k_cache 0.00140159 0.80674417 - layer.3.v_cache 0.00000237 0.00270936 - layer.4.k_cache 0.00336527 1.59354490 - layer.4.v_cache 0.00000311 0.00426807 - layer.4.output 0.00016660 0.11013498 - ------------------------------------------------------------------------------------- - TOTAL 0.00250283 1.95718561 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 316984 -BPFP 1.1486 bits/point -EBPFP 2.2973 equivalent bits/point -MSE 1.957186 ----------------------- -------------------------------------------------------- -Time: 0.652s Load: 0.008s, Pack+Encode: 0.259s, Decode+Unpack: 0.385s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9572 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample13-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample13-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample130-layer4-item1.zst (10/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample130-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 100, 128) -Output shape: (1, 100, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.output: torch.Size([1, 100, 4096]) -> torch.Size([1, 1, 100, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,080B, BPFP=0.2406 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,200B, BPFP=1.5781 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,556B, BPFP=0.9028 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 21,632B, BPFP=1.6900 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,152B, BPFP=1.1838 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 21,904B, BPFP=1.7112 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,616B, BPFP=1.1419 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,764B, BPFP=1.7003 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,264B, BPFP=0.8800 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,332B, BPFP=1.7447 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 70,208B, BPFP=1.3713 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02583573 25.85609863 - layer.0.v_cache 0.00000027 0.00031400 - layer.1.k_cache 0.00330333 1.64037994 - layer.1.v_cache 0.00000088 0.00115118 - layer.2.k_cache 0.00114776 0.74683784 - layer.2.v_cache 0.00000114 0.00162620 - layer.3.k_cache 0.00139474 0.86730591 - layer.3.v_cache 0.00000229 0.00279115 - layer.4.k_cache 0.00328125 1.76025070 - layer.4.v_cache 0.00000316 0.00455934 - layer.4.output 0.00020624 0.13720753 - ------------------------------------------------------------------------------------- - TOTAL 0.00255682 2.24501036 - (elements=1,433,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1433600 -Total Bytes 233708 -BPFP 1.3042 bits/point -EBPFP 2.6083 equivalent bits/point -MSE 2.245010 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.007s, Pack+Encode: 0.208s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2450 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample130-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample130-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample144-layer4-item1.zst (11/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample144-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,844B, BPFP=0.2339 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,804B, BPFP=1.6286 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,524B, BPFP=0.8655 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,264B, BPFP=1.6664 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,000B, BPFP=1.0691 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,208B, BPFP=1.6618 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,268B, BPFP=1.0911 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,584B, BPFP=1.6105 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,804B, BPFP=0.8885 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,512B, BPFP=1.6868 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 50,728B, BPFP=1.0429 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.287s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02533972 23.79125977 - layer.0.v_cache 0.00000026 0.00031007 - layer.1.k_cache 0.00345234 1.69547537 - layer.1.v_cache 0.00000076 0.00109175 - layer.2.k_cache 0.00126989 0.76512162 - layer.2.v_cache 0.00000106 0.00153899 - layer.3.k_cache 0.00140666 0.83536650 - layer.3.v_cache 0.00000216 0.00269579 - layer.4.k_cache 0.00351815 1.67769406 - layer.4.v_cache 0.00000309 0.00462814 - layer.4.output 0.00017514 0.14436886 - ------------------------------------------------------------------------------------- - TOTAL 0.00254962 2.09661839 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 201540 -BPFP 1.1839 bits/point -EBPFP 2.3677 equivalent bits/point -MSE 2.096618 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.008s, Pack+Encode: 0.206s, Decode+Unpack: 0.287s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0966 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample144-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample144-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample145-layer4-item1.zst (12/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample145-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,712B, BPFP=0.2184 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,048B, BPFP=1.5341 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,740B, BPFP=0.7845 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,792B, BPFP=1.5941 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,636B, BPFP=1.0177 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 19,988B, BPFP=1.6099 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,948B, BPFP=1.0428 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,992B, BPFP=1.6102 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,700B, BPFP=0.7812 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,696B, BPFP=1.6669 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,188B, BPFP=0.7689 -⌛️ [2/4] FRONTEND: Frontend time: 0.205s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02570412 24.25433916 - layer.0.v_cache 0.00000027 0.00030361 - layer.1.k_cache 0.00344234 1.68458683 - layer.1.v_cache 0.00000085 0.00107314 - layer.2.k_cache 0.00116154 0.76969485 - layer.2.v_cache 0.00000108 0.00154729 - layer.3.k_cache 0.00150460 0.84379821 - layer.3.v_cache 0.00000231 0.00258779 - layer.4.k_cache 0.00334272 1.73273177 - layer.4.v_cache 0.00000310 0.00440784 - layer.4.output 0.00020164 0.13265376 - ------------------------------------------------------------------------------------- - TOTAL 0.00256925 2.13040611 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 185440 -BPFP 1.0668 bits/point -EBPFP 2.1337 equivalent bits/point -MSE 2.130406 ----------------------- -------------------------------------------------------- -Time: 0.499s Load: 0.006s, Pack+Encode: 0.205s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1304 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample145-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample145-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample146-layer4-item1.zst (13/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample146-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 98, 128) -Output shape: (1, 98, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.output: torch.Size([1, 98, 4096]) -> torch.Size([1, 1, 98, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,844B, BPFP=0.2267 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,356B, BPFP=1.5430 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,292B, BPFP=0.8205 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,320B, BPFP=1.6199 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,252B, BPFP=1.1362 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,880B, BPFP=1.6645 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,696B, BPFP=1.0918 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,868B, BPFP=1.6636 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,156B, BPFP=0.8893 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,504B, BPFP=1.7143 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 67,640B, BPFP=1.3481 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02660144 24.44705138 - layer.0.v_cache 0.00000028 0.00030458 - layer.1.k_cache 0.00330389 1.60754348 - layer.1.v_cache 0.00000100 0.00112723 - layer.2.k_cache 0.00113731 0.75124826 - layer.2.v_cache 0.00000122 0.00167461 - layer.3.k_cache 0.00133429 0.85939438 - layer.3.v_cache 0.00000236 0.00278171 - layer.4.k_cache 0.00329518 1.70518696 - layer.4.v_cache 0.00000318 0.00451426 - layer.4.output 0.00020743 0.12798356 - ------------------------------------------------------------------------------------- - TOTAL 0.00260785 2.13519722 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 222808 -BPFP 1.2687 bits/point -EBPFP 2.5374 equivalent bits/point -MSE 2.135197 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.006s, Pack+Encode: 0.208s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1352 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample146-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample146-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample147-layer4-item1.zst (14/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample147-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 98, 128) -Output shape: (1, 98, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.output: torch.Size([1, 98, 4096]) -> torch.Size([1, 1, 98, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,820B, BPFP=0.2248 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,388B, BPFP=1.5456 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,164B, BPFP=0.8103 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,296B, BPFP=1.6180 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,660B, BPFP=1.0890 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,684B, BPFP=1.6489 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,580B, BPFP=1.0826 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,016B, BPFP=1.6754 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,448B, BPFP=0.8329 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,172B, BPFP=1.6878 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 52,552B, BPFP=1.0474 -⌛️ [2/4] FRONTEND: Frontend time: 0.207s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02655566 23.90655891 - layer.0.v_cache 0.00000028 0.00032107 - layer.1.k_cache 0.00343622 1.60852393 - layer.1.v_cache 0.00000088 0.00116794 - layer.2.k_cache 0.00122497 0.75211996 - layer.2.v_cache 0.00000114 0.00165253 - layer.3.k_cache 0.00135168 0.83960560 - layer.3.v_cache 0.00000224 0.00275344 - layer.4.k_cache 0.00342672 1.67636358 - layer.4.v_cache 0.00000300 0.00438667 - layer.4.output 0.00022158 0.12577738 - ------------------------------------------------------------------------------------- - TOTAL 0.00263494 2.09261165 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 205780 -BPFP 1.1718 bits/point -EBPFP 2.3435 equivalent bits/point -MSE 2.092612 ----------------------- -------------------------------------------------------- -Time: 0.502s Load: 0.007s, Pack+Encode: 0.207s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0926 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample147-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample147-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample150-layer4-item1.zst (15/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample150-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 100, 128) -Output shape: (1, 100, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.output: torch.Size([1, 100, 4096]) -> torch.Size([1, 1, 100, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,028B, BPFP=0.2366 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,220B, BPFP=1.5797 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,428B, BPFP=0.8928 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 21,164B, BPFP=1.6534 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,036B, BPFP=1.0966 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 21,932B, BPFP=1.7134 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,784B, BPFP=1.1550 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,664B, BPFP=1.6925 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,044B, BPFP=0.8628 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,904B, BPFP=1.7112 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 50,772B, BPFP=0.9916 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02664840 24.67058838 - layer.0.v_cache 0.00000027 0.00030692 - layer.1.k_cache 0.00335783 1.64277496 - layer.1.v_cache 0.00000086 0.00108678 - layer.2.k_cache 0.00109031 0.73964798 - layer.2.v_cache 0.00000115 0.00156537 - layer.3.k_cache 0.00136642 0.84346298 - layer.3.v_cache 0.00000222 0.00263523 - layer.4.k_cache 0.00341075 1.68671890 - layer.4.v_cache 0.00000317 0.00429766 - layer.4.output 0.00031254 0.12521528 - ------------------------------------------------------------------------------------- - TOTAL 0.00265226 2.14956759 - (elements=1,433,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1433600 -Total Bytes 211976 -BPFP 1.1829 bits/point -EBPFP 2.3658 equivalent bits/point -MSE 2.149568 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.006s, Pack+Encode: 0.206s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1496 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample150-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample150-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample153-layer4-item1.zst (16/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample153-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,700B, BPFP=0.2175 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,656B, BPFP=1.5831 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,776B, BPFP=0.7874 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,284B, BPFP=1.6337 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,868B, BPFP=1.0364 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,384B, BPFP=1.6418 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,160B, BPFP=1.0599 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,136B, BPFP=1.6218 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,572B, BPFP=0.7709 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,952B, BPFP=1.6875 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,824B, BPFP=0.9025 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.290s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02654818 23.88551818 - layer.0.v_cache 0.00000027 0.00032348 - layer.1.k_cache 0.00347710 1.69146760 - layer.1.v_cache 0.00000089 0.00121681 - layer.2.k_cache 0.00114081 0.76557875 - layer.2.v_cache 0.00000111 0.00166170 - layer.3.k_cache 0.00140715 0.87178795 - layer.3.v_cache 0.00000211 0.00276533 - layer.4.k_cache 0.00324570 1.79127109 - layer.4.v_cache 0.00000329 0.00468384 - layer.4.output 0.00023085 0.13659408 - ------------------------------------------------------------------------------------- - TOTAL 0.00262500 2.11161793 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 194312 -BPFP 1.1179 bits/point -EBPFP 2.2357 equivalent bits/point -MSE 2.111618 ----------------------- -------------------------------------------------------- -Time: 0.506s Load: 0.007s, Pack+Encode: 0.209s, Decode+Unpack: 0.290s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1116 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample153-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample153-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample154-layer4-item1.zst (17/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample154-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,732B, BPFP=0.2398 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,168B, BPFP=1.5948 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,080B, BPFP=0.7971 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,784B, BPFP=1.6489 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,720B, BPFP=1.0288 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 19,720B, BPFP=1.7310 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,504B, BPFP=1.0976 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,484B, BPFP=1.7103 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,952B, BPFP=0.7858 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,924B, BPFP=1.7489 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 36,188B, BPFP=0.7942 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02697376 28.14074021 - layer.0.v_cache 0.00000027 0.00029703 - layer.1.k_cache 0.00348241 1.81697065 - layer.1.v_cache 0.00000086 0.00107558 - layer.2.k_cache 0.00112258 0.76315891 - layer.2.v_cache 0.00000113 0.00155449 - layer.3.k_cache 0.00136851 0.88295146 - layer.3.v_cache 0.00000223 0.00252826 - layer.4.k_cache 0.00328884 1.78824470 - layer.4.v_cache 0.00000300 0.00415023 - layer.4.output 0.00019035 0.14796986 - ------------------------------------------------------------------------------------- - TOTAL 0.00264321 2.42811078 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 177256 -BPFP 1.1114 bits/point -EBPFP 2.2228 equivalent bits/point -MSE 2.428111 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.006s, Pack+Encode: 0.206s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4281 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample154-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample154-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample157-layer4-item1.zst (18/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample157-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 104, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 104, 128) -Output shape: (1, 104, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.0.v_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.1.k_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.1.v_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.2.k_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.2.v_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.3.k_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.3.v_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.4.k_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.4.v_cache: torch.Size([1, 8, 104, 128]) -> torch.Size([1, 1, 104, 1024]) - layer.4.output: torch.Size([1, 104, 4096]) -> torch.Size([1, 1, 104, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,108B, BPFP=0.2335 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,064B, BPFP=1.6575 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,472B, BPFP=0.8618 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,816B, BPFP=1.7139 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,204B, BPFP=1.0670 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,564B, BPFP=1.6950 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 15,112B, BPFP=1.1352 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,184B, BPFP=1.6665 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,872B, BPFP=0.8918 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,004B, BPFP=1.7281 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 67,856B, BPFP=1.2743 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 104, 128]) - layer.0.v_cache: torch.Size([1, 8, 104, 128]) - layer.1.k_cache: torch.Size([1, 8, 104, 128]) - layer.1.v_cache: torch.Size([1, 8, 104, 128]) - layer.2.k_cache: torch.Size([1, 8, 104, 128]) - layer.2.v_cache: torch.Size([1, 8, 104, 128]) - layer.3.k_cache: torch.Size([1, 8, 104, 128]) - layer.3.v_cache: torch.Size([1, 8, 104, 128]) - layer.4.k_cache: torch.Size([1, 8, 104, 128]) - layer.4.v_cache: torch.Size([1, 8, 104, 128]) - layer.4.output: torch.Size([1, 104, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 104, 128]) - layer.0.v_cache: torch.Size([1, 8, 104, 128]) - layer.1.k_cache: torch.Size([1, 8, 104, 128]) - layer.1.v_cache: torch.Size([1, 8, 104, 128]) - layer.2.k_cache: torch.Size([1, 8, 104, 128]) - layer.2.v_cache: torch.Size([1, 8, 104, 128]) - layer.3.k_cache: torch.Size([1, 8, 104, 128]) - layer.3.v_cache: torch.Size([1, 8, 104, 128]) - layer.4.k_cache: torch.Size([1, 8, 104, 128]) - layer.4.v_cache: torch.Size([1, 8, 104, 128]) - layer.4.output: torch.Size([1, 104, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02512770 25.91450383 - layer.0.v_cache 0.00000027 0.00032291 - layer.1.k_cache 0.00323274 1.69163997 - layer.1.v_cache 0.00000093 0.00118038 - layer.2.k_cache 0.00116098 0.73146585 - layer.2.v_cache 0.00000116 0.00168680 - layer.3.k_cache 0.00137668 0.82831361 - layer.3.v_cache 0.00000213 0.00273726 - layer.4.k_cache 0.00339989 1.68171575 - layer.4.v_cache 0.00000305 0.00446121 - layer.4.output 0.00020049 0.12977770 - ------------------------------------------------------------------------------------- - TOTAL 0.00250768 2.24122417 - (elements=1,490,944) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1490944 -Total Bytes 236256 -BPFP 1.2677 bits/point -EBPFP 2.5354 equivalent bits/point -MSE 2.241224 ----------------------- -------------------------------------------------------- -Time: 0.504s Load: 0.007s, Pack+Encode: 0.208s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 104, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 104, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2412 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample157-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample157-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample159-layer4-item1.zst (19/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample159-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 82, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 82, 128) -Output shape: (1, 82, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.output: torch.Size([1, 82, 4096]) -> torch.Size([1, 1, 82, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,624B, BPFP=0.2500 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 15,864B, BPFP=1.5114 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,404B, BPFP=0.8007 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,252B, BPFP=1.6437 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,824B, BPFP=1.0312 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 17,960B, BPFP=1.7111 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,344B, BPFP=1.0808 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,584B, BPFP=1.6753 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,028B, BPFP=0.7649 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 18,300B, BPFP=1.7435 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 31,780B, BPFP=0.7570 -⌛️ [2/4] FRONTEND: Frontend time: 0.207s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.output: torch.Size([1, 82, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.output: torch.Size([1, 82, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02805163 29.72669052 - layer.0.v_cache 0.00000027 0.00030373 - layer.1.k_cache 0.00355746 1.96923493 - layer.1.v_cache 0.00000072 0.00103557 - layer.2.k_cache 0.00112546 0.79250075 - layer.2.v_cache 0.00000104 0.00150143 - layer.3.k_cache 0.00143209 0.90118278 - layer.3.v_cache 0.00000205 0.00251787 - layer.4.k_cache 0.00323843 1.87175472 - layer.4.v_cache 0.00000284 0.00402395 - layer.4.output 0.00022045 0.14882561 - ------------------------------------------------------------------------------------- - TOTAL 0.00273527 2.56186062 - (elements=1,175,552) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1175552 -Total Bytes 159964 -BPFP 1.0886 bits/point -EBPFP 2.1772 equivalent bits/point -MSE 2.561861 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.006s, Pack+Encode: 0.207s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 82, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5619 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample159-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample159-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample165-layer4-item1.zst (20/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample165-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 109, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 109, 128) -Output shape: (1, 109, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.0.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.1.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.1.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.2.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.2.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.3.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.3.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.4.k_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.4.v_cache: torch.Size([1, 8, 109, 128]) -> torch.Size([1, 1, 109, 1024]) - layer.4.output: torch.Size([1, 109, 4096]) -> torch.Size([1, 1, 109, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,116B, BPFP=0.2233 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,888B, BPFP=1.5688 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,096B, BPFP=0.7953 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,428B, BPFP=1.6075 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,732B, BPFP=0.9842 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,444B, BPFP=1.6087 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,172B, BPFP=1.0158 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,588B, BPFP=1.5473 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,996B, BPFP=0.8598 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,876B, BPFP=1.6396 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 55,136B, BPFP=0.9880 -⌛️ [2/4] FRONTEND: Frontend time: 0.207s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 109, 128]) - layer.0.v_cache: torch.Size([1, 8, 109, 128]) - layer.1.k_cache: torch.Size([1, 8, 109, 128]) - layer.1.v_cache: torch.Size([1, 8, 109, 128]) - layer.2.k_cache: torch.Size([1, 8, 109, 128]) - layer.2.v_cache: torch.Size([1, 8, 109, 128]) - layer.3.k_cache: torch.Size([1, 8, 109, 128]) - layer.3.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.k_cache: torch.Size([1, 8, 109, 128]) - layer.4.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.output: torch.Size([1, 109, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 109, 128]) - layer.0.v_cache: torch.Size([1, 8, 109, 128]) - layer.1.k_cache: torch.Size([1, 8, 109, 128]) - layer.1.v_cache: torch.Size([1, 8, 109, 128]) - layer.2.k_cache: torch.Size([1, 8, 109, 128]) - layer.2.v_cache: torch.Size([1, 8, 109, 128]) - layer.3.k_cache: torch.Size([1, 8, 109, 128]) - layer.3.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.k_cache: torch.Size([1, 8, 109, 128]) - layer.4.v_cache: torch.Size([1, 8, 109, 128]) - layer.4.output: torch.Size([1, 109, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02442908 23.73067929 - layer.0.v_cache 0.00000028 0.00031700 - layer.1.k_cache 0.00329736 1.62177914 - layer.1.v_cache 0.00000085 0.00111233 - layer.2.k_cache 0.00114487 0.73091531 - layer.2.v_cache 0.00000123 0.00163654 - layer.3.k_cache 0.00134342 0.82862931 - layer.3.v_cache 0.00000206 0.00262968 - layer.4.k_cache 0.00348010 1.70530771 - layer.4.v_cache 0.00000308 0.00436053 - layer.4.output 0.00017639 0.12206130 - ------------------------------------------------------------------------------------- - TOTAL 0.00245771 2.07968657 - (elements=1,562,624) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1562624 -Total Bytes 220472 -BPFP 1.1287 bits/point -EBPFP 2.2575 equivalent bits/point -MSE 2.079687 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.007s, Pack+Encode: 0.207s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 109, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 109, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0797 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample165-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample165-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample167-layer4-item1.zst (21/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample167-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,732B, BPFP=0.2247 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,756B, BPFP=1.6247 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,084B, BPFP=0.9115 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,760B, BPFP=1.7072 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,184B, BPFP=1.0842 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,852B, BPFP=1.7148 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,856B, BPFP=1.1395 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,936B, BPFP=1.7217 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,412B, BPFP=0.8562 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,704B, BPFP=1.7849 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 57,676B, BPFP=1.1858 -⌛️ [2/4] FRONTEND: Frontend time: 0.211s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.287s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02733303 23.12460424 - layer.0.v_cache 0.00000028 0.00031937 - layer.1.k_cache 0.00367247 1.67024504 - layer.1.v_cache 0.00000082 0.00114440 - layer.2.k_cache 0.00112740 0.74267361 - layer.2.v_cache 0.00000120 0.00171100 - layer.3.k_cache 0.00137809 0.85370403 - layer.3.v_cache 0.00000227 0.00276022 - layer.4.k_cache 0.00319271 1.65544707 - layer.4.v_cache 0.00000332 0.00454134 - layer.4.output 0.00022674 0.12652290 - ------------------------------------------------------------------------------------- - TOTAL 0.00268704 2.04023156 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 212952 -BPFP 1.2509 bits/point -EBPFP 2.5018 equivalent bits/point -MSE 2.040232 ----------------------- -------------------------------------------------------- -Time: 0.504s Load: 0.006s, Pack+Encode: 0.211s, Decode+Unpack: 0.287s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0402 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample167-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample167-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample168-layer4-item1.zst (22/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample168-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 103, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 103, 128) -Output shape: (1, 103, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.output: torch.Size([1, 103, 4096]) -> torch.Size([1, 1, 103, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,076B, BPFP=0.2333 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,744B, BPFP=1.6493 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,444B, BPFP=0.8680 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,536B, BPFP=1.7093 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,160B, BPFP=1.1499 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,828B, BPFP=1.7315 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 15,080B, BPFP=1.1438 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,964B, BPFP=1.6660 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,844B, BPFP=0.8984 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,008B, BPFP=1.7451 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 50,372B, BPFP=0.9552 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.output: torch.Size([1, 103, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.292s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.output: torch.Size([1, 103, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03002706 26.03664480 - layer.0.v_cache 0.00000028 0.00030429 - layer.1.k_cache 0.00329593 1.61330051 - layer.1.v_cache 0.00000087 0.00114941 - layer.2.k_cache 0.00114743 0.72651006 - layer.2.v_cache 0.00000126 0.00159132 - layer.3.k_cache 0.00133553 0.82037939 - layer.3.v_cache 0.00000217 0.00265863 - layer.4.k_cache 0.00338515 1.62655254 - layer.4.v_cache 0.00000318 0.00440819 - layer.4.output 0.00017408 0.12137260 - ------------------------------------------------------------------------------------- - TOTAL 0.00284966 2.23707068 - (elements=1,476,608) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1476608 -Total Bytes 219056 -BPFP 1.1868 bits/point -EBPFP 2.3736 equivalent bits/point -MSE 2.237071 ----------------------- -------------------------------------------------------- -Time: 0.505s Load: 0.007s, Pack+Encode: 0.206s, Decode+Unpack: 0.292s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 103, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2371 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample168-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample168-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample174-layer4-item1.zst (23/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample174-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 103, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 103, 128) -Output shape: (1, 103, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.output: torch.Size([1, 103, 4096]) -> torch.Size([1, 1, 103, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,080B, BPFP=0.2336 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,732B, BPFP=1.6484 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,776B, BPFP=0.8932 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,204B, BPFP=1.6842 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,296B, BPFP=1.1602 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,480B, BPFP=1.7051 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 15,048B, BPFP=1.1414 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,704B, BPFP=1.7221 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,356B, BPFP=0.9372 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,992B, BPFP=1.7439 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 63,784B, BPFP=1.2095 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.output: torch.Size([1, 103, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.291s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.output: torch.Size([1, 103, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02636531 25.73378243 - layer.0.v_cache 0.00000028 0.00031177 - layer.1.k_cache 0.00340973 1.64400934 - layer.1.v_cache 0.00000088 0.00110394 - layer.2.k_cache 0.00113950 0.74606531 - layer.2.v_cache 0.00000120 0.00162628 - layer.3.k_cache 0.00133761 0.84706664 - layer.3.v_cache 0.00000227 0.00268352 - layer.4.k_cache 0.00339299 1.66971388 - layer.4.v_cache 0.00000316 0.00444876 - layer.4.output 0.00023648 0.13172160 - ------------------------------------------------------------------------------------- - TOTAL 0.00261420 2.22697845 - (elements=1,476,608) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1476608 -Total Bytes 233452 -BPFP 1.2648 bits/point -EBPFP 2.5296 equivalent bits/point -MSE 2.226978 ----------------------- -------------------------------------------------------- -Time: 0.507s Load: 0.007s, Pack+Encode: 0.209s, Decode+Unpack: 0.291s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 103, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2270 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample174-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample174-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample182-layer4-item1.zst (24/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample182-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,860B, BPFP=0.2429 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,964B, BPFP=1.6104 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,640B, BPFP=0.8186 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,400B, BPFP=1.7323 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,052B, BPFP=1.0234 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,616B, BPFP=1.7507 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,008B, BPFP=1.1046 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,716B, BPFP=1.7592 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,004B, BPFP=0.8495 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,468B, BPFP=1.8230 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 49,704B, BPFP=1.0552 -⌛️ [2/4] FRONTEND: Frontend time: 0.210s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02458628 26.64447021 - layer.0.v_cache 0.00000028 0.00031265 - layer.1.k_cache 0.00347829 1.87104913 - layer.1.v_cache 0.00000086 0.00113236 - layer.2.k_cache 0.00113727 0.73585991 - layer.2.v_cache 0.00000116 0.00158671 - layer.3.k_cache 0.00132770 0.86047206 - layer.3.v_cache 0.00000243 0.00271799 - layer.4.k_cache 0.00339865 1.76376011 - layer.4.v_cache 0.00000297 0.00424603 - layer.4.output 0.00018871 0.13195997 - ------------------------------------------------------------------------------------- - TOTAL 0.00247791 2.31524622 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 199432 -BPFP 1.2097 bits/point -EBPFP 2.4194 equivalent bits/point -MSE 2.315246 ----------------------- -------------------------------------------------------- -Time: 0.505s Load: 0.005s, Pack+Encode: 0.210s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.3152 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample182-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample182-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample185-layer4-item1.zst (25/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample185-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,692B, BPFP=0.2191 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,224B, BPFP=1.5645 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,736B, BPFP=0.7923 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,692B, BPFP=1.6025 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,848B, BPFP=1.0456 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,244B, BPFP=1.6475 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,956B, BPFP=1.0544 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,880B, BPFP=1.6178 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,540B, BPFP=0.7764 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,660B, BPFP=1.6813 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 52,420B, BPFP=1.0665 -⌛️ [2/4] FRONTEND: Frontend time: 0.207s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02634023 23.59547933 - layer.0.v_cache 0.00000027 0.00029758 - layer.1.k_cache 0.00351230 1.58963251 - layer.1.v_cache 0.00000075 0.00102671 - layer.2.k_cache 0.00114531 0.71292003 - layer.2.v_cache 0.00000110 0.00150461 - layer.3.k_cache 0.00136477 0.82199240 - layer.3.v_cache 0.00000217 0.00261087 - layer.4.k_cache 0.00356451 1.66943232 - layer.4.v_cache 0.00000291 0.00408251 - layer.4.output 0.00018809 0.13457675 - ------------------------------------------------------------------------------------- - TOTAL 0.00262048 2.06694899 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 199892 -BPFP 1.1619 bits/point -EBPFP 2.3239 equivalent bits/point -MSE 2.066949 ----------------------- -------------------------------------------------------- -Time: 0.504s Load: 0.008s, Pack+Encode: 0.207s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0669 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample185-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample185-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample191-layer4-item1.zst (26/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample191-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,948B, BPFP=0.2476 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,312B, BPFP=1.6223 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,772B, BPFP=0.9049 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,100B, BPFP=1.6885 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,812B, BPFP=1.0763 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,856B, BPFP=1.7520 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,428B, BPFP=1.1280 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,688B, BPFP=1.7379 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,296B, BPFP=0.8649 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,416B, BPFP=1.7991 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 47,872B, BPFP=1.0054 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02618607 24.41364247 - layer.0.v_cache 0.00000028 0.00030847 - layer.1.k_cache 0.00358459 1.78592280 - layer.1.v_cache 0.00000078 0.00110437 - layer.2.k_cache 0.00113483 0.76280541 - layer.2.v_cache 0.00000114 0.00157735 - layer.3.k_cache 0.00136506 0.83268410 - layer.3.v_cache 0.00000214 0.00264487 - layer.4.k_cache 0.00334622 1.69996692 - layer.4.v_cache 0.00000303 0.00430032 - layer.4.output 0.00022383 0.14535787 - ------------------------------------------------------------------------------------- - TOTAL 0.00260854 2.14902775 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 200500 -BPFP 1.2031 bits/point -EBPFP 2.4062 equivalent bits/point -MSE 2.149028 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.006s, Pack+Encode: 0.209s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1490 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample191-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample191-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample196-layer4-item1.zst (27/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample196-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,760B, BPFP=0.2246 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,732B, BPFP=1.6058 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,292B, BPFP=0.8376 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,804B, BPFP=1.6930 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,300B, BPFP=1.0824 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 21,012B, BPFP=1.7100 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,228B, BPFP=1.0765 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,724B, BPFP=1.6865 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,908B, BPFP=0.8063 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,288B, BPFP=1.7324 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 57,512B, BPFP=1.1701 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02641018 22.59703573 - layer.0.v_cache 0.00000028 0.00030606 - layer.1.k_cache 0.00334712 1.60575199 - layer.1.v_cache 0.00000079 0.00114159 - layer.2.k_cache 0.00118939 0.74002767 - layer.2.v_cache 0.00000116 0.00169912 - layer.3.k_cache 0.00138481 0.82363733 - layer.3.v_cache 0.00000224 0.00277372 - layer.4.k_cache 0.00336247 1.55815617 - layer.4.v_cache 0.00000319 0.00448984 - layer.4.output 0.00019453 0.11918860 - ------------------------------------------------------------------------------------- - TOTAL 0.00260570 1.98655526 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 210560 -BPFP 1.2240 bits/point -EBPFP 2.4479 equivalent bits/point -MSE 1.986555 ----------------------- -------------------------------------------------------- -Time: 0.504s Load: 0.007s, Pack+Encode: 0.208s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9866 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample196-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample196-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample197-layer4-item1.zst (28/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample197-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,856B, BPFP=0.2425 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,700B, BPFP=1.5880 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,024B, BPFP=0.8512 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,228B, BPFP=1.7177 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,456B, BPFP=1.1427 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 21,296B, BPFP=1.8084 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,916B, BPFP=1.1817 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,536B, BPFP=1.7439 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,884B, BPFP=0.8393 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,928B, BPFP=1.7772 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,228B, BPFP=0.9177 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.293s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02542763 26.42470385 - layer.0.v_cache 0.00000028 0.00030552 - layer.1.k_cache 0.00330708 1.80221790 - layer.1.v_cache 0.00000075 0.00112594 - layer.2.k_cache 0.00113337 0.77306184 - layer.2.v_cache 0.00000121 0.00171152 - layer.3.k_cache 0.00137180 0.86363826 - layer.3.v_cache 0.00000212 0.00275114 - layer.4.k_cache 0.00335567 1.73878628 - layer.4.v_cache 0.00000315 0.00446217 - layer.4.output 0.00020664 0.13374101 - ------------------------------------------------------------------------------------- - TOTAL 0.00253069 2.29626632 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 195052 -BPFP 1.1831 bits/point -EBPFP 2.3662 equivalent bits/point -MSE 2.296266 ----------------------- -------------------------------------------------------- -Time: 0.506s Load: 0.006s, Pack+Encode: 0.208s, Decode+Unpack: 0.293s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2963 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample197-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample197-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample2-layer4-item1.zst (29/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample2-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 178, 128) -Output shape: (1, 178, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.output: torch.Size([1, 178, 4096]) -> torch.Size([1, 1, 178, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,908B, BPFP=0.2154 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 30,972B, BPFP=1.3594 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 15,024B, BPFP=0.6594 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 32,928B, BPFP=1.4452 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,320B, BPFP=0.8041 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 33,132B, BPFP=1.4542 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,948B, BPFP=0.8755 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 32,176B, BPFP=1.4122 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 14,892B, BPFP=0.6536 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 33,448B, BPFP=1.4680 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 86,528B, BPFP=0.9494 -⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.378s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02409235 21.41408307 - layer.0.v_cache 0.00000026 0.00029525 - layer.1.k_cache 0.00309449 1.43054036 - layer.1.v_cache 0.00000082 0.00105089 - layer.2.k_cache 0.00118451 0.70858229 - layer.2.v_cache 0.00000118 0.00153097 - layer.3.k_cache 0.00136933 0.79859933 - layer.3.v_cache 0.00000226 0.00264427 - layer.4.k_cache 0.00340187 1.57448706 - layer.4.v_cache 0.00000300 0.00412975 - layer.4.output 0.00017401 0.10968573 - ------------------------------------------------------------------------------------- - TOTAL 0.00241758 1.88390615 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 322276 -BPFP 1.0103 bits/point -EBPFP 2.0207 equivalent bits/point -MSE 1.883906 ----------------------- -------------------------------------------------------- -Time: 0.644s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.378s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8839 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample2-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample2-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample201-layer4-item1.zst (30/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample201-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,716B, BPFP=0.2234 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,492B, BPFP=1.6030 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,496B, BPFP=0.8632 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,884B, BPFP=1.7174 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,664B, BPFP=1.1237 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,860B, BPFP=1.7155 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,736B, BPFP=1.1296 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,612B, BPFP=1.6951 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,172B, BPFP=0.9187 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,132B, BPFP=1.7378 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 54,636B, BPFP=1.1233 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.281s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02560628 22.48620991 - layer.0.v_cache 0.00000028 0.00030072 - layer.1.k_cache 0.00332260 1.71350033 - layer.1.v_cache 0.00000077 0.00109933 - layer.2.k_cache 0.00114542 0.75348190 - layer.2.v_cache 0.00000120 0.00156467 - layer.3.k_cache 0.00143781 0.86144714 - layer.3.v_cache 0.00000216 0.00252235 - layer.4.k_cache 0.00324414 1.74520344 - layer.4.v_cache 0.00000298 0.00405791 - layer.4.output 0.00027467 0.14652396 - ------------------------------------------------------------------------------------- - TOTAL 0.00256159 2.01110597 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 209400 -BPFP 1.2300 bits/point -EBPFP 2.4601 equivalent bits/point -MSE 2.011106 ----------------------- -------------------------------------------------------- -Time: 0.493s Load: 0.006s, Pack+Encode: 0.206s, Decode+Unpack: 0.281s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0111 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample201-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample201-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample212-layer4-item1.zst (31/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample212-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,844B, BPFP=0.2364 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,056B, BPFP=1.6669 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,812B, BPFP=0.8986 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,712B, BPFP=1.7214 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,416B, BPFP=1.1150 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,760B, BPFP=1.7254 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,836B, BPFP=1.1499 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,248B, BPFP=1.6828 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,732B, BPFP=0.8920 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,640B, BPFP=1.7985 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 49,452B, BPFP=1.0275 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02535593 23.92126984 - layer.0.v_cache 0.00000029 0.00029607 - layer.1.k_cache 0.00347189 1.64739698 - layer.1.v_cache 0.00000073 0.00099716 - layer.2.k_cache 0.00115370 0.73931779 - layer.2.v_cache 0.00000109 0.00151936 - layer.3.k_cache 0.00136415 0.82365012 - layer.3.v_cache 0.00000218 0.00252804 - layer.4.k_cache 0.00350239 1.72945810 - layer.4.v_cache 0.00000301 0.00423203 - layer.4.output 0.00017689 0.12996089 - ------------------------------------------------------------------------------------- - TOTAL 0.00254021 2.09932208 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 204508 -BPFP 1.2141 bits/point -EBPFP 2.4281 equivalent bits/point -MSE 2.099322 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.006s, Pack+Encode: 0.206s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0993 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample212-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample212-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample214-layer4-item1.zst (32/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample214-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 99, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 99, 128) -Output shape: (1, 99, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.0.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.1.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.1.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.2.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.2.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.3.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.3.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.4.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.4.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.4.output: torch.Size([1, 99, 4096]) -> torch.Size([1, 1, 99, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,956B, BPFP=0.2333 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,756B, BPFP=1.5590 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,236B, BPFP=0.8867 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 21,092B, BPFP=1.6645 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,404B, BPFP=1.1367 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,896B, BPFP=1.6490 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,676B, BPFP=1.0792 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,324B, BPFP=1.6039 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,832B, BPFP=0.8548 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,684B, BPFP=1.7112 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 48,976B, BPFP=0.9662 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 99, 128]) - layer.0.v_cache: torch.Size([1, 8, 99, 128]) - layer.1.k_cache: torch.Size([1, 8, 99, 128]) - layer.1.v_cache: torch.Size([1, 8, 99, 128]) - layer.2.k_cache: torch.Size([1, 8, 99, 128]) - layer.2.v_cache: torch.Size([1, 8, 99, 128]) - layer.3.k_cache: torch.Size([1, 8, 99, 128]) - layer.3.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.k_cache: torch.Size([1, 8, 99, 128]) - layer.4.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.output: torch.Size([1, 99, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.284s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 99, 128]) - layer.0.v_cache: torch.Size([1, 8, 99, 128]) - layer.1.k_cache: torch.Size([1, 8, 99, 128]) - layer.1.v_cache: torch.Size([1, 8, 99, 128]) - layer.2.k_cache: torch.Size([1, 8, 99, 128]) - layer.2.v_cache: torch.Size([1, 8, 99, 128]) - layer.3.k_cache: torch.Size([1, 8, 99, 128]) - layer.3.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.k_cache: torch.Size([1, 8, 99, 128]) - layer.4.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.output: torch.Size([1, 99, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02440176 23.36391010 - layer.0.v_cache 0.00000028 0.00030236 - layer.1.k_cache 0.00326434 1.68499201 - layer.1.v_cache 0.00000081 0.00107723 - layer.2.k_cache 0.00115982 0.74944390 - layer.2.v_cache 0.00000131 0.00163414 - layer.3.k_cache 0.00139121 0.85499241 - layer.3.v_cache 0.00000205 0.00261765 - layer.4.k_cache 0.00335893 1.62023278 - layer.4.v_cache 0.00000299 0.00435672 - layer.4.output 0.00020438 0.12210490 - ------------------------------------------------------------------------------------- - TOTAL 0.00245722 2.05514135 - (elements=1,419,264) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1419264 -Total Bytes 205832 -BPFP 1.1602 bits/point -EBPFP 2.3204 equivalent bits/point -MSE 2.055141 ----------------------- -------------------------------------------------------- -Time: 0.495s Load: 0.007s, Pack+Encode: 0.204s, Decode+Unpack: 0.284s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 99, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0551 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample214-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample214-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample224-layer4-item1.zst (33/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample224-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 103, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 103, 128) -Output shape: (1, 103, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) -> torch.Size([1, 1, 103, 1024]) - layer.4.output: torch.Size([1, 103, 4096]) -> torch.Size([1, 1, 103, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,060B, BPFP=0.2321 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,320B, BPFP=1.6171 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,588B, BPFP=0.8789 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,140B, BPFP=1.6793 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,068B, BPFP=1.1429 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,436B, BPFP=1.7018 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 15,044B, BPFP=1.1411 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,320B, BPFP=1.6930 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,244B, BPFP=0.9287 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,444B, BPFP=1.7024 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 63,672B, BPFP=1.2074 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.output: torch.Size([1, 103, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.292s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 103, 128]) - layer.0.v_cache: torch.Size([1, 8, 103, 128]) - layer.1.k_cache: torch.Size([1, 8, 103, 128]) - layer.1.v_cache: torch.Size([1, 8, 103, 128]) - layer.2.k_cache: torch.Size([1, 8, 103, 128]) - layer.2.v_cache: torch.Size([1, 8, 103, 128]) - layer.3.k_cache: torch.Size([1, 8, 103, 128]) - layer.3.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.k_cache: torch.Size([1, 8, 103, 128]) - layer.4.v_cache: torch.Size([1, 8, 103, 128]) - layer.4.output: torch.Size([1, 103, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02603974 27.40295055 - layer.0.v_cache 0.00000026 0.00029242 - layer.1.k_cache 0.00325535 1.67649101 - layer.1.v_cache 0.00000093 0.00109784 - layer.2.k_cache 0.00112553 0.74679995 - layer.2.v_cache 0.00000114 0.00153227 - layer.3.k_cache 0.00141218 0.85740210 - layer.3.v_cache 0.00000204 0.00253644 - layer.4.k_cache 0.00324713 1.75142884 - layer.4.v_cache 0.00000302 0.00419843 - layer.4.output 0.00020093 0.13804836 - ------------------------------------------------------------------------------------- - TOTAL 0.00256365 2.35692309 - (elements=1,476,608) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1476608 -Total Bytes 231336 -BPFP 1.2533 bits/point -EBPFP 2.5067 equivalent bits/point -MSE 2.356923 ----------------------- -------------------------------------------------------- -Time: 0.508s Load: 0.007s, Pack+Encode: 0.209s, Decode+Unpack: 0.292s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 103, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 103, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.3569 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample224-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample224-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample227-layer4-item1.zst (34/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample227-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,772B, BPFP=0.2256 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,876B, BPFP=1.6175 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,912B, BPFP=0.8880 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,644B, BPFP=1.6800 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,540B, BPFP=1.1019 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,688B, BPFP=1.6836 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,788B, BPFP=1.1221 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,208B, BPFP=1.6445 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,620B, BPFP=0.8643 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,224B, BPFP=1.7272 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 54,436B, BPFP=1.1075 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.293s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02591230 22.40774790 - layer.0.v_cache 0.00000028 0.00029265 - layer.1.k_cache 0.00321835 1.55836836 - layer.1.v_cache 0.00000081 0.00115127 - layer.2.k_cache 0.00115341 0.74671133 - layer.2.v_cache 0.00000115 0.00157567 - layer.3.k_cache 0.00138578 0.82655303 - layer.3.v_cache 0.00000239 0.00276627 - layer.4.k_cache 0.00338578 1.64405187 - layer.4.v_cache 0.00000322 0.00448888 - layer.4.output 0.00018605 0.11521860 - ------------------------------------------------------------------------------------- - TOTAL 0.00255769 1.97532726 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 208708 -BPFP 1.2132 bits/point -EBPFP 2.4264 equivalent bits/point -MSE 1.975327 ----------------------- -------------------------------------------------------- -Time: 0.507s Load: 0.006s, Pack+Encode: 0.208s, Decode+Unpack: 0.293s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9753 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample227-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample227-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample233-layer4-item1.zst (35/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample233-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,744B, BPFP=0.2522 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 17,072B, BPFP=1.5691 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,772B, BPFP=0.8982 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,620B, BPFP=1.7114 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,736B, BPFP=1.1706 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 19,092B, BPFP=1.7548 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,532B, BPFP=1.1518 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 18,672B, BPFP=1.7162 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,248B, BPFP=0.8500 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,280B, BPFP=1.7721 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 49,236B, BPFP=1.1313 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.292s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02725663 28.11738281 - layer.0.v_cache 0.00000027 0.00032144 - layer.1.k_cache 0.00328410 1.70720269 - layer.1.v_cache 0.00000085 0.00119196 - layer.2.k_cache 0.00114159 0.78839578 - layer.2.v_cache 0.00000125 0.00179931 - layer.3.k_cache 0.00133976 0.88708397 - layer.3.v_cache 0.00000266 0.00298312 - layer.4.k_cache 0.00332377 1.74180854 - layer.4.v_cache 0.00000336 0.00488135 - layer.4.output 0.00023926 0.13980754 - ------------------------------------------------------------------------------------- - TOTAL 0.00266509 2.41516294 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 189004 -BPFP 1.2408 bits/point -EBPFP 2.4817 equivalent bits/point -MSE 2.415163 ----------------------- -------------------------------------------------------- -Time: 0.506s Load: 0.005s, Pack+Encode: 0.209s, Decode+Unpack: 0.292s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4152 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample233-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample233-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample241-layer4-item1.zst (36/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample241-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,668B, BPFP=0.2149 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,220B, BPFP=1.5480 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,184B, BPFP=0.8202 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,952B, BPFP=1.6070 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,916B, BPFP=1.0403 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,228B, BPFP=1.6292 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,144B, BPFP=1.0586 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,916B, BPFP=1.6041 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,204B, BPFP=0.8218 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,456B, BPFP=1.6476 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 47,176B, BPFP=0.9499 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.293s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02553191 24.71890101 - layer.0.v_cache 0.00000029 0.00032628 - layer.1.k_cache 0.00322990 1.67785896 - layer.1.v_cache 0.00000089 0.00113585 - layer.2.k_cache 0.00117012 0.74683884 - layer.2.v_cache 0.00000108 0.00161641 - layer.3.k_cache 0.00135684 0.84326093 - layer.3.v_cache 0.00000209 0.00260547 - layer.4.k_cache 0.00331928 1.71679719 - layer.4.v_cache 0.00000299 0.00424743 - layer.4.output 0.00021425 0.13969156 - ------------------------------------------------------------------------------------- - TOTAL 0.00253374 2.16231104 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 196064 -BPFP 1.1279 bits/point -EBPFP 2.2559 equivalent bits/point -MSE 2.162311 ----------------------- -------------------------------------------------------- -Time: 0.508s Load: 0.006s, Pack+Encode: 0.209s, Decode+Unpack: 0.293s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1623 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample241-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample241-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample250-layer4-item1.zst (37/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample250-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,712B, BPFP=0.2184 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,072B, BPFP=1.6166 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,596B, BPFP=0.8534 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,576B, BPFP=1.6572 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,364B, BPFP=1.0764 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,804B, BPFP=1.6756 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,568B, BPFP=1.0928 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,784B, BPFP=1.6740 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,008B, BPFP=0.8866 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,568B, BPFP=1.7371 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 62,300B, BPFP=1.2544 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.293s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02408928 24.96995560 - layer.0.v_cache 0.00000028 0.00032401 - layer.1.k_cache 0.00330002 1.70629631 - layer.1.v_cache 0.00000090 0.00113635 - layer.2.k_cache 0.00113900 0.74976632 - layer.2.v_cache 0.00000130 0.00164202 - layer.3.k_cache 0.00136322 0.84883999 - layer.3.v_cache 0.00000251 0.00287151 - layer.4.k_cache 0.00342626 1.62288768 - layer.4.v_cache 0.00000319 0.00444290 - layer.4.output 0.00021399 0.13630620 - ------------------------------------------------------------------------------------- - TOTAL 0.00244156 2.17524196 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 217352 -BPFP 1.2504 bits/point -EBPFP 2.5008 equivalent bits/point -MSE 2.175242 ----------------------- -------------------------------------------------------- -Time: 0.507s Load: 0.007s, Pack+Encode: 0.208s, Decode+Unpack: 0.293s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1752 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample250-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample250-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample251-layer4-item1.zst (38/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample251-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 99, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 99, 128) -Output shape: (1, 99, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.0.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.1.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.1.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.2.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.2.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.3.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.3.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.4.k_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.4.v_cache: torch.Size([1, 8, 99, 128]) -> torch.Size([1, 1, 99, 1024]) - layer.4.output: torch.Size([1, 99, 4096]) -> torch.Size([1, 1, 99, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,896B, BPFP=0.2285 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,148B, BPFP=1.5900 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,468B, BPFP=0.9050 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,468B, BPFP=1.6152 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,556B, BPFP=1.1487 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,468B, BPFP=1.6152 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,828B, BPFP=1.0912 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,444B, BPFP=1.6133 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,528B, BPFP=0.8308 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,044B, BPFP=1.7396 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 47,252B, BPFP=0.9322 -⌛️ [2/4] FRONTEND: Frontend time: 0.210s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 99, 128]) - layer.0.v_cache: torch.Size([1, 8, 99, 128]) - layer.1.k_cache: torch.Size([1, 8, 99, 128]) - layer.1.v_cache: torch.Size([1, 8, 99, 128]) - layer.2.k_cache: torch.Size([1, 8, 99, 128]) - layer.2.v_cache: torch.Size([1, 8, 99, 128]) - layer.3.k_cache: torch.Size([1, 8, 99, 128]) - layer.3.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.k_cache: torch.Size([1, 8, 99, 128]) - layer.4.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.output: torch.Size([1, 99, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.292s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 99, 128]) - layer.0.v_cache: torch.Size([1, 8, 99, 128]) - layer.1.k_cache: torch.Size([1, 8, 99, 128]) - layer.1.v_cache: torch.Size([1, 8, 99, 128]) - layer.2.k_cache: torch.Size([1, 8, 99, 128]) - layer.2.v_cache: torch.Size([1, 8, 99, 128]) - layer.3.k_cache: torch.Size([1, 8, 99, 128]) - layer.3.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.k_cache: torch.Size([1, 8, 99, 128]) - layer.4.v_cache: torch.Size([1, 8, 99, 128]) - layer.4.output: torch.Size([1, 99, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02490362 25.14600596 - layer.0.v_cache 0.00000029 0.00031701 - layer.1.k_cache 0.00335574 1.59370314 - layer.1.v_cache 0.00000081 0.00103560 - layer.2.k_cache 0.00112037 0.74134380 - layer.2.v_cache 0.00000110 0.00148151 - layer.3.k_cache 0.00138480 0.83668202 - layer.3.v_cache 0.00000194 0.00243524 - layer.4.k_cache 0.00346640 1.69871937 - layer.4.v_cache 0.00000296 0.00399904 - layer.4.output 0.00019387 0.11974488 - ------------------------------------------------------------------------------------- - TOTAL 0.00250096 2.17890730 - (elements=1,419,264) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1419264 -Total Bytes 204100 -BPFP 1.1505 bits/point -EBPFP 2.3009 equivalent bits/point -MSE 2.178907 ----------------------- -------------------------------------------------------- -Time: 0.509s Load: 0.008s, Pack+Encode: 0.210s, Decode+Unpack: 0.292s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 99, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 99, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1789 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample251-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample251-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample257-layer4-item1.zst (39/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample257-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 86, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 86, 128) -Output shape: (1, 86, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.0.v_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.1.k_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.1.v_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.2.k_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.2.v_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.3.k_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.3.v_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.4.k_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.4.v_cache: torch.Size([1, 8, 86, 128]) -> torch.Size([1, 1, 86, 1024]) - layer.4.output: torch.Size([1, 86, 4096]) -> torch.Size([1, 1, 86, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,776B, BPFP=0.2522 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,180B, BPFP=1.6515 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,644B, BPFP=0.8761 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,084B, BPFP=1.7336 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,524B, BPFP=1.1377 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,132B, BPFP=1.8289 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,152B, BPFP=1.1948 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,652B, BPFP=1.7852 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,004B, BPFP=0.8180 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,964B, BPFP=1.8136 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,444B, BPFP=1.0548 -⌛️ [2/4] FRONTEND: Frontend time: 0.210s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 86, 128]) - layer.0.v_cache: torch.Size([1, 8, 86, 128]) - layer.1.k_cache: torch.Size([1, 8, 86, 128]) - layer.1.v_cache: torch.Size([1, 8, 86, 128]) - layer.2.k_cache: torch.Size([1, 8, 86, 128]) - layer.2.v_cache: torch.Size([1, 8, 86, 128]) - layer.3.k_cache: torch.Size([1, 8, 86, 128]) - layer.3.v_cache: torch.Size([1, 8, 86, 128]) - layer.4.k_cache: torch.Size([1, 8, 86, 128]) - layer.4.v_cache: torch.Size([1, 8, 86, 128]) - layer.4.output: torch.Size([1, 86, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.291s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 86, 128]) - layer.0.v_cache: torch.Size([1, 8, 86, 128]) - layer.1.k_cache: torch.Size([1, 8, 86, 128]) - layer.1.v_cache: torch.Size([1, 8, 86, 128]) - layer.2.k_cache: torch.Size([1, 8, 86, 128]) - layer.2.v_cache: torch.Size([1, 8, 86, 128]) - layer.3.k_cache: torch.Size([1, 8, 86, 128]) - layer.3.v_cache: torch.Size([1, 8, 86, 128]) - layer.4.k_cache: torch.Size([1, 8, 86, 128]) - layer.4.v_cache: torch.Size([1, 8, 86, 128]) - layer.4.output: torch.Size([1, 86, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02609516 25.48495696 - layer.0.v_cache 0.00000027 0.00030062 - layer.1.k_cache 0.00348465 1.72887492 - layer.1.v_cache 0.00000080 0.00113935 - layer.2.k_cache 0.00118276 0.77527175 - layer.2.v_cache 0.00000133 0.00182833 - layer.3.k_cache 0.00137135 0.89803953 - layer.3.v_cache 0.00000258 0.00295202 - layer.4.k_cache 0.00338397 1.76073899 - layer.4.v_cache 0.00000307 0.00458332 - layer.4.output 0.00019889 0.13605535 - ------------------------------------------------------------------------------------- - TOTAL 0.00259439 2.22877909 - (elements=1,232,896) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1232896 -Total Bytes 190556 -BPFP 1.2365 bits/point -EBPFP 2.4730 equivalent bits/point -MSE 2.228779 ----------------------- -------------------------------------------------------- -Time: 0.507s Load: 0.006s, Pack+Encode: 0.210s, Decode+Unpack: 0.291s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 86, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 86, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2288 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample257-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample257-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample258-layer4-item1.zst (40/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample258-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,784B, BPFP=0.2444 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 17,352B, BPFP=1.5232 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,416B, BPFP=0.8265 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,460B, BPFP=1.6204 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,964B, BPFP=1.0502 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 19,620B, BPFP=1.7223 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,816B, BPFP=1.1250 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,100B, BPFP=1.6766 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,668B, BPFP=0.7609 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,776B, BPFP=1.7360 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 45,396B, BPFP=0.9962 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02681794 26.94219079 - layer.0.v_cache 0.00000027 0.00031635 - layer.1.k_cache 0.00344681 1.79368883 - layer.1.v_cache 0.00000079 0.00114938 - layer.2.k_cache 0.00114729 0.77930699 - layer.2.v_cache 0.00000121 0.00166898 - layer.3.k_cache 0.00138721 0.88861744 - layer.3.v_cache 0.00000216 0.00281451 - layer.4.k_cache 0.00341441 1.89367419 - layer.4.v_cache 0.00000316 0.00467294 - layer.4.output 0.00022172 0.14823695 - ------------------------------------------------------------------------------------- - TOTAL 0.00265058 2.35007487 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 185352 -BPFP 1.1622 bits/point -EBPFP 2.3243 equivalent bits/point -MSE 2.350075 ----------------------- -------------------------------------------------------- -Time: 0.502s Load: 0.006s, Pack+Encode: 0.206s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.3501 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample258-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample258-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample263-layer4-item1.zst (41/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample263-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,756B, BPFP=0.2419 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,020B, BPFP=1.5818 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,576B, BPFP=0.8406 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,996B, BPFP=1.7553 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,584B, BPFP=1.1046 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,420B, BPFP=1.7925 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,568B, BPFP=1.1032 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,960B, BPFP=1.7521 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,460B, BPFP=0.8304 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,268B, BPFP=1.7791 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 47,932B, BPFP=1.0519 -⌛️ [2/4] FRONTEND: Frontend time: 0.273s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.281s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02424761 25.23419944 - layer.0.v_cache 0.00000029 0.00031148 - layer.1.k_cache 0.00343373 1.71689940 - layer.1.v_cache 0.00000078 0.00115137 - layer.2.k_cache 0.00114421 0.73810157 - layer.2.v_cache 0.00000115 0.00164714 - layer.3.k_cache 0.00135741 0.84544638 - layer.3.v_cache 0.00000236 0.00271906 - layer.4.k_cache 0.00332665 1.67267926 - layer.4.v_cache 0.00000319 0.00447490 - layer.4.output 0.00018046 0.12339721 - ------------------------------------------------------------------------------------- - TOTAL 0.00244566 2.19365849 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 193540 -BPFP 1.2135 bits/point -EBPFP 2.4270 equivalent bits/point -MSE 2.193658 ----------------------- -------------------------------------------------------- -Time: 0.561s Load: 0.007s, Pack+Encode: 0.273s, Decode+Unpack: 0.281s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1937 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample263-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample263-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample274-layer4-item1.zst (42/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample274-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,828B, BPFP=0.2401 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,452B, BPFP=1.5669 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,692B, BPFP=0.9079 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,864B, BPFP=1.6868 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,856B, BPFP=1.0917 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,616B, BPFP=1.7507 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,860B, BPFP=1.1770 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,244B, BPFP=1.7191 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,580B, BPFP=0.8135 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,816B, BPFP=1.7677 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,304B, BPFP=0.8132 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.280s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02551763 25.38998280 - layer.0.v_cache 0.00000027 0.00030145 - layer.1.k_cache 0.00347722 1.77536641 - layer.1.v_cache 0.00000076 0.00101022 - layer.2.k_cache 0.00117662 0.76020730 - layer.2.v_cache 0.00000108 0.00151255 - layer.3.k_cache 0.00141157 0.84925992 - layer.3.v_cache 0.00000195 0.00241489 - layer.4.k_cache 0.00338827 1.73009076 - layer.4.v_cache 0.00000294 0.00424103 - layer.4.output 0.00021021 0.13642560 - ------------------------------------------------------------------------------------- - TOTAL 0.00255851 2.21857784 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 188112 -BPFP 1.1410 bits/point -EBPFP 2.2820 equivalent bits/point -MSE 2.218578 ----------------------- -------------------------------------------------------- -Time: 0.485s Load: 0.005s, Pack+Encode: 0.200s, Decode+Unpack: 0.280s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2186 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample274-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample274-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample282-layer4-item1.zst (43/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample282-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,756B, BPFP=0.2475 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 17,588B, BPFP=1.5794 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,568B, BPFP=0.8592 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,812B, BPFP=1.6893 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,544B, BPFP=1.1264 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 19,508B, BPFP=1.7518 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,872B, BPFP=1.1559 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,292B, BPFP=1.7324 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,032B, BPFP=0.8111 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,292B, BPFP=1.7324 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,888B, BPFP=0.9179 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02591714 23.83867075 - layer.0.v_cache 0.00000026 0.00029872 - layer.1.k_cache 0.00374749 1.72877871 - layer.1.v_cache 0.00000077 0.00109343 - layer.2.k_cache 0.00116905 0.76353876 - layer.2.v_cache 0.00000114 0.00159456 - layer.3.k_cache 0.00135864 0.84352542 - layer.3.v_cache 0.00000233 0.00274975 - layer.4.k_cache 0.00339113 1.71383667 - layer.4.v_cache 0.00000304 0.00430241 - layer.4.output 0.00030438 0.13707608 - ------------------------------------------------------------------------------------- - TOTAL 0.00262918 2.10333525 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 182152 -BPFP 1.1684 bits/point -EBPFP 2.3367 equivalent bits/point -MSE 2.103335 ----------------------- -------------------------------------------------------- -Time: 0.482s Load: 0.006s, Pack+Encode: 0.199s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1033 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample282-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample282-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample290-layer4-item1.zst (44/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample290-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,836B, BPFP=0.2357 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,992B, BPFP=1.6616 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,492B, BPFP=0.8720 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,292B, BPFP=1.6865 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,596B, BPFP=1.1300 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,296B, BPFP=1.6868 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,928B, BPFP=1.1576 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,840B, BPFP=1.6489 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,080B, BPFP=0.8378 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,796B, BPFP=1.7284 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 52,540B, BPFP=1.0917 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02423150 24.79490369 - layer.0.v_cache 0.00000028 0.00030163 - layer.1.k_cache 0.00344140 1.76000960 - layer.1.v_cache 0.00000074 0.00105535 - layer.2.k_cache 0.00110852 0.73842734 - layer.2.v_cache 0.00000113 0.00156572 - layer.3.k_cache 0.00135732 0.83397545 - layer.3.v_cache 0.00000223 0.00268511 - layer.4.k_cache 0.00337875 1.59619449 - layer.4.v_cache 0.00000288 0.00411511 - layer.4.output 0.00018208 0.11807960 - ------------------------------------------------------------------------------------- - TOTAL 0.00244665 2.15753942 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 204688 -BPFP 1.2151 bits/point -EBPFP 2.4303 equivalent bits/point -MSE 2.157539 ----------------------- -------------------------------------------------------- -Time: 0.483s Load: 0.007s, Pack+Encode: 0.200s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1575 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample290-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample290-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample3-layer4-item1.zst (45/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample3-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 149, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 149, 128) -Output shape: (1, 149, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.0.v_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.1.k_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.1.v_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.2.k_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.2.v_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.3.k_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.3.v_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.4.k_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.4.v_cache: torch.Size([1, 8, 149, 128]) -> torch.Size([1, 1, 149, 1024]) - layer.4.output: torch.Size([1, 149, 4096]) -> torch.Size([1, 1, 149, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,948B, BPFP=0.2594 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 29,288B, BPFP=1.5357 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,724B, BPFP=0.7720 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 31,352B, BPFP=1.6439 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,280B, BPFP=0.9585 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 31,744B, BPFP=1.6644 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,768B, BPFP=1.0365 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 30,976B, BPFP=1.6242 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 14,964B, BPFP=0.7846 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 31,996B, BPFP=1.6776 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 86,448B, BPFP=1.1332 -⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 149, 128]) - layer.0.v_cache: torch.Size([1, 8, 149, 128]) - layer.1.k_cache: torch.Size([1, 8, 149, 128]) - layer.1.v_cache: torch.Size([1, 8, 149, 128]) - layer.2.k_cache: torch.Size([1, 8, 149, 128]) - layer.2.v_cache: torch.Size([1, 8, 149, 128]) - layer.3.k_cache: torch.Size([1, 8, 149, 128]) - layer.3.v_cache: torch.Size([1, 8, 149, 128]) - layer.4.k_cache: torch.Size([1, 8, 149, 128]) - layer.4.v_cache: torch.Size([1, 8, 149, 128]) - layer.4.output: torch.Size([1, 149, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.377s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 149, 128]) - layer.0.v_cache: torch.Size([1, 8, 149, 128]) - layer.1.k_cache: torch.Size([1, 8, 149, 128]) - layer.1.v_cache: torch.Size([1, 8, 149, 128]) - layer.2.k_cache: torch.Size([1, 8, 149, 128]) - layer.2.v_cache: torch.Size([1, 8, 149, 128]) - layer.3.k_cache: torch.Size([1, 8, 149, 128]) - layer.3.v_cache: torch.Size([1, 8, 149, 128]) - layer.4.k_cache: torch.Size([1, 8, 149, 128]) - layer.4.v_cache: torch.Size([1, 8, 149, 128]) - layer.4.output: torch.Size([1, 149, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02473939 21.23537450 - layer.0.v_cache 0.00000028 0.00032187 - layer.1.k_cache 0.00314236 1.40020363 - layer.1.v_cache 0.00000089 0.00116173 - layer.2.k_cache 0.00115430 0.72661801 - layer.2.v_cache 0.00000109 0.00163915 - layer.3.k_cache 0.00134690 0.81057376 - layer.3.v_cache 0.00000207 0.00271092 - layer.4.k_cache 0.00351049 1.52180368 - layer.4.v_cache 0.00000318 0.00460221 - layer.4.output 0.00015300 0.10439168 - ------------------------------------------------------------------------------------- - TOTAL 0.00246521 1.86589830 - (elements=2,136,064) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2136064 -Total Bytes 314488 -BPFP 1.1778 bits/point -EBPFP 2.3556 equivalent bits/point -MSE 1.865898 ----------------------- -------------------------------------------------------- -Time: 0.634s Load: 0.009s, Pack+Encode: 0.248s, Decode+Unpack: 0.377s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 149, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 149, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8659 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample3-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample3-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample307-layer4-item1.zst (46/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample307-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,908B, BPFP=0.2443 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 19,276B, BPFP=1.6193 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,792B, BPFP=0.9066 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,692B, BPFP=1.7382 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,372B, BPFP=1.1233 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 21,116B, BPFP=1.7739 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,392B, BPFP=1.1250 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,040B, BPFP=1.7675 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,032B, BPFP=0.9267 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 21,536B, BPFP=1.8091 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 54,192B, BPFP=1.1381 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.278s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02417885 24.20865360 - layer.0.v_cache 0.00000029 0.00030886 - layer.1.k_cache 0.00337012 1.74355111 - layer.1.v_cache 0.00000087 0.00107531 - layer.2.k_cache 0.00111931 0.73761692 - layer.2.v_cache 0.00000117 0.00150469 - layer.3.k_cache 0.00133665 0.82107987 - layer.3.v_cache 0.00000243 0.00259029 - layer.4.k_cache 0.00331438 1.65296099 - layer.4.v_cache 0.00000316 0.00429318 - layer.4.output 0.00021472 0.13063790 - ------------------------------------------------------------------------------------- - TOTAL 0.00244186 2.12115617 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 209348 -BPFP 1.2562 bits/point -EBPFP 2.5123 equivalent bits/point -MSE 2.121156 ----------------------- -------------------------------------------------------- -Time: 0.483s Load: 0.005s, Pack+Encode: 0.200s, Decode+Unpack: 0.278s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1212 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample307-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample307-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample313-layer4-item1.zst (47/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample313-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 81, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 81, 128) -Output shape: (1, 81, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.0.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.1.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.1.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.2.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.2.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.3.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.3.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.4.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.4.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.4.output: torch.Size([1, 81, 4096]) -> torch.Size([1, 1, 81, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,620B, BPFP=0.2527 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 17,200B, BPFP=1.6590 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,016B, BPFP=0.8696 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,320B, BPFP=1.7670 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,148B, BPFP=1.1717 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 19,700B, BPFP=1.9001 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,044B, BPFP=1.1617 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 19,320B, BPFP=1.8634 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,344B, BPFP=0.8048 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,292B, BPFP=1.8607 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,968B, BPFP=1.0120 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 81, 128]) - layer.0.v_cache: torch.Size([1, 8, 81, 128]) - layer.1.k_cache: torch.Size([1, 8, 81, 128]) - layer.1.v_cache: torch.Size([1, 8, 81, 128]) - layer.2.k_cache: torch.Size([1, 8, 81, 128]) - layer.2.v_cache: torch.Size([1, 8, 81, 128]) - layer.3.k_cache: torch.Size([1, 8, 81, 128]) - layer.3.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.k_cache: torch.Size([1, 8, 81, 128]) - layer.4.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.output: torch.Size([1, 81, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.278s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 81, 128]) - layer.0.v_cache: torch.Size([1, 8, 81, 128]) - layer.1.k_cache: torch.Size([1, 8, 81, 128]) - layer.1.v_cache: torch.Size([1, 8, 81, 128]) - layer.2.k_cache: torch.Size([1, 8, 81, 128]) - layer.2.v_cache: torch.Size([1, 8, 81, 128]) - layer.3.k_cache: torch.Size([1, 8, 81, 128]) - layer.3.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.k_cache: torch.Size([1, 8, 81, 128]) - layer.4.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.output: torch.Size([1, 81, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02598516 28.95078306 - layer.0.v_cache 0.00000026 0.00031075 - layer.1.k_cache 0.00346748 1.87322471 - layer.1.v_cache 0.00000092 0.00115432 - layer.2.k_cache 0.00113221 0.78268993 - layer.2.v_cache 0.00000143 0.00168825 - layer.3.k_cache 0.00133539 0.87130370 - layer.3.v_cache 0.00000229 0.00277330 - layer.4.k_cache 0.00342405 1.79438876 - layer.4.v_cache 0.00000310 0.00441151 - layer.4.output 0.00020696 0.13175060 - ------------------------------------------------------------------------------------- - TOTAL 0.00258429 2.48640934 - (elements=1,161,216) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1161216 -Total Bytes 179972 -BPFP 1.2399 bits/point -EBPFP 2.4798 equivalent bits/point -MSE 2.486409 ----------------------- -------------------------------------------------------- -Time: 0.482s Load: 0.005s, Pack+Encode: 0.199s, Decode+Unpack: 0.278s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 81, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4864 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample313-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample313-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample319-layer4-item1.zst (48/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample319-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,928B, BPFP=0.2460 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,620B, BPFP=1.5642 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,788B, BPFP=0.9062 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,824B, BPFP=1.6653 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,232B, BPFP=1.0276 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,444B, BPFP=1.7174 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,772B, BPFP=1.0729 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,136B, BPFP=1.6915 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,972B, BPFP=0.8377 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,800B, BPFP=1.7473 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,840B, BPFP=0.8577 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02450277 25.70874023 - layer.0.v_cache 0.00000028 0.00031367 - layer.1.k_cache 0.00336978 1.81033194 - layer.1.v_cache 0.00000078 0.00110528 - layer.2.k_cache 0.00117610 0.74463170 - layer.2.v_cache 0.00000111 0.00161006 - layer.3.k_cache 0.00133121 0.84424509 - layer.3.v_cache 0.00000213 0.00262712 - layer.4.k_cache 0.00346230 1.73478666 - layer.4.v_cache 0.00000315 0.00438100 - layer.4.output 0.00018648 0.12975907 - ------------------------------------------------------------------------------------- - TOTAL 0.00247111 2.24084350 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 189356 -BPFP 1.1362 bits/point -EBPFP 2.2724 equivalent bits/point -MSE 2.240844 ----------------------- -------------------------------------------------------- -Time: 0.483s Load: 0.006s, Pack+Encode: 0.200s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2408 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample319-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample319-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample333-layer4-item1.zst (49/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample333-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,784B, BPFP=0.2417 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 18,572B, BPFP=1.6122 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,752B, BPFP=0.8465 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 19,308B, BPFP=1.6760 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,308B, BPFP=1.0684 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,160B, BPFP=1.7500 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,168B, BPFP=1.1431 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,304B, BPFP=1.7625 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,104B, BPFP=0.7903 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,048B, BPFP=1.7403 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 35,888B, BPFP=0.7788 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02906363 27.91430935 - layer.0.v_cache 0.00000028 0.00030208 - layer.1.k_cache 0.00333766 1.85870090 - layer.1.v_cache 0.00000075 0.00102598 - layer.2.k_cache 0.00114903 0.74398651 - layer.2.v_cache 0.00000112 0.00145192 - layer.3.k_cache 0.00131913 0.87562917 - layer.3.v_cache 0.00000216 0.00247863 - layer.4.k_cache 0.00329656 1.72683529 - layer.4.v_cache 0.00000300 0.00407709 - layer.4.output 0.00024019 0.13198553 - ------------------------------------------------------------------------------------- - TOTAL 0.00279529 2.40405279 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 181396 -BPFP 1.1247 bits/point -EBPFP 2.2495 equivalent bits/point -MSE 2.404053 ----------------------- -------------------------------------------------------- -Time: 0.481s Load: 0.005s, Pack+Encode: 0.199s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4041 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample333-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample333-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample36-layer4-item1.zst (50/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample36-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 119, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 119, 128) -Output shape: (1, 119, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.0.v_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.1.k_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.1.v_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.2.k_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.2.v_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.3.k_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.3.v_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.4.k_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.4.v_cache: torch.Size([1, 8, 119, 128]) -> torch.Size([1, 1, 119, 1024]) - layer.4.output: torch.Size([1, 119, 4096]) -> torch.Size([1, 1, 119, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,288B, BPFP=0.2159 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,244B, BPFP=1.3947 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,492B, BPFP=0.6888 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,060B, BPFP=1.5139 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,224B, BPFP=0.8682 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 23,336B, BPFP=1.5320 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,856B, BPFP=0.9097 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,284B, BPFP=1.4630 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,416B, BPFP=0.6838 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,856B, BPFP=1.5662 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 66,628B, BPFP=1.0936 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 119, 128]) - layer.0.v_cache: torch.Size([1, 8, 119, 128]) - layer.1.k_cache: torch.Size([1, 8, 119, 128]) - layer.1.v_cache: torch.Size([1, 8, 119, 128]) - layer.2.k_cache: torch.Size([1, 8, 119, 128]) - layer.2.v_cache: torch.Size([1, 8, 119, 128]) - layer.3.k_cache: torch.Size([1, 8, 119, 128]) - layer.3.v_cache: torch.Size([1, 8, 119, 128]) - layer.4.k_cache: torch.Size([1, 8, 119, 128]) - layer.4.v_cache: torch.Size([1, 8, 119, 128]) - layer.4.output: torch.Size([1, 119, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.280s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 119, 128]) - layer.0.v_cache: torch.Size([1, 8, 119, 128]) - layer.1.k_cache: torch.Size([1, 8, 119, 128]) - layer.1.v_cache: torch.Size([1, 8, 119, 128]) - layer.2.k_cache: torch.Size([1, 8, 119, 128]) - layer.2.v_cache: torch.Size([1, 8, 119, 128]) - layer.3.k_cache: torch.Size([1, 8, 119, 128]) - layer.3.v_cache: torch.Size([1, 8, 119, 128]) - layer.4.k_cache: torch.Size([1, 8, 119, 128]) - layer.4.v_cache: torch.Size([1, 8, 119, 128]) - layer.4.output: torch.Size([1, 119, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02458241 20.96427546 - layer.0.v_cache 0.00000028 0.00030700 - layer.1.k_cache 0.00331476 1.39061872 - layer.1.v_cache 0.00000080 0.00115366 - layer.2.k_cache 0.00114230 0.70589165 - layer.2.v_cache 0.00000115 0.00171656 - layer.3.k_cache 0.00138709 0.80838513 - layer.3.v_cache 0.00000225 0.00282656 - layer.4.k_cache 0.00333808 1.59406838 - layer.4.v_cache 0.00000314 0.00442252 - layer.4.output 0.00020645 0.11908200 - ------------------------------------------------------------------------------------- - TOTAL 0.00247129 1.85357097 - (elements=1,705,984) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1705984 -Total Bytes 231684 -BPFP 1.0865 bits/point -EBPFP 2.1729 equivalent bits/point -MSE 1.853571 ----------------------- -------------------------------------------------------- -Time: 0.487s Load: 0.007s, Pack+Encode: 0.199s, Decode+Unpack: 0.280s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 119, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 119, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8536 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample36-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample36-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample365-layer4-item1.zst (51/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample365-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 82, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 82, 128) -Output shape: (1, 82, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.output: torch.Size([1, 82, 4096]) -> torch.Size([1, 1, 82, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,612B, BPFP=0.2489 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 16,476B, BPFP=1.5697 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,440B, BPFP=0.8041 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,404B, BPFP=1.6582 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,636B, BPFP=1.0133 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 18,420B, BPFP=1.7550 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,468B, BPFP=1.0926 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,800B, BPFP=1.6959 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,856B, BPFP=0.7485 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 18,164B, BPFP=1.7306 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 33,696B, BPFP=0.8026 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.output: torch.Size([1, 82, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.output: torch.Size([1, 82, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02610582 29.87281464 - layer.0.v_cache 0.00000027 0.00031624 - layer.1.k_cache 0.00344214 1.83914092 - layer.1.v_cache 0.00000078 0.00111789 - layer.2.k_cache 0.00114699 0.78328072 - layer.2.v_cache 0.00000119 0.00166608 - layer.3.k_cache 0.00140111 0.87544846 - layer.3.v_cache 0.00000200 0.00264461 - layer.4.k_cache 0.00332288 1.75752314 - layer.4.v_cache 0.00000299 0.00419207 - layer.4.output 0.00018752 0.13442871 - ------------------------------------------------------------------------------------- - TOTAL 0.00258402 2.54827569 - (elements=1,175,552) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1175552 -Total Bytes 162972 -BPFP 1.1091 bits/point -EBPFP 2.2182 equivalent bits/point -MSE 2.548276 ----------------------- -------------------------------------------------------- -Time: 0.482s Load: 0.006s, Pack+Encode: 0.200s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 82, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5483 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample365-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample365-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample38-layer4-item1.zst (52/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample38-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 124, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 124, 128) -Output shape: (1, 124, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.0.v_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.1.k_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.1.v_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.2.k_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.2.v_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.3.k_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.3.v_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.4.k_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.4.v_cache: torch.Size([1, 8, 124, 128]) -> torch.Size([1, 1, 124, 1024]) - layer.4.output: torch.Size([1, 124, 4096]) -> torch.Size([1, 1, 124, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,392B, BPFP=0.2137 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,988B, BPFP=1.3853 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,740B, BPFP=0.6767 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,236B, BPFP=1.5270 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,580B, BPFP=0.8556 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 24,052B, BPFP=1.5154 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,628B, BPFP=0.9216 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,536B, BPFP=1.4199 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,564B, BPFP=0.6656 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 24,152B, BPFP=1.5217 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 58,852B, BPFP=0.9270 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 124, 128]) - layer.0.v_cache: torch.Size([1, 8, 124, 128]) - layer.1.k_cache: torch.Size([1, 8, 124, 128]) - layer.1.v_cache: torch.Size([1, 8, 124, 128]) - layer.2.k_cache: torch.Size([1, 8, 124, 128]) - layer.2.v_cache: torch.Size([1, 8, 124, 128]) - layer.3.k_cache: torch.Size([1, 8, 124, 128]) - layer.3.v_cache: torch.Size([1, 8, 124, 128]) - layer.4.k_cache: torch.Size([1, 8, 124, 128]) - layer.4.v_cache: torch.Size([1, 8, 124, 128]) - layer.4.output: torch.Size([1, 124, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.281s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 124, 128]) - layer.0.v_cache: torch.Size([1, 8, 124, 128]) - layer.1.k_cache: torch.Size([1, 8, 124, 128]) - layer.1.v_cache: torch.Size([1, 8, 124, 128]) - layer.2.k_cache: torch.Size([1, 8, 124, 128]) - layer.2.v_cache: torch.Size([1, 8, 124, 128]) - layer.3.k_cache: torch.Size([1, 8, 124, 128]) - layer.3.v_cache: torch.Size([1, 8, 124, 128]) - layer.4.k_cache: torch.Size([1, 8, 124, 128]) - layer.4.v_cache: torch.Size([1, 8, 124, 128]) - layer.4.output: torch.Size([1, 124, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02574280 19.59957000 - layer.0.v_cache 0.00000027 0.00030105 - layer.1.k_cache 0.00324291 1.41719178 - layer.1.v_cache 0.00000080 0.00105171 - layer.2.k_cache 0.00117169 0.71443109 - layer.2.v_cache 0.00000108 0.00149471 - layer.3.k_cache 0.00137525 0.79619038 - layer.3.v_cache 0.00000203 0.00243824 - layer.4.k_cache 0.00340243 1.58408786 - layer.4.v_cache 0.00000299 0.00409214 - layer.4.output 0.00019780 0.11860456 - ------------------------------------------------------------------------------------- - TOTAL 0.00255239 1.75680480 - (elements=1,777,664) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1777664 -Total Bytes 228720 -BPFP 1.0293 bits/point -EBPFP 2.0586 equivalent bits/point -MSE 1.756805 ----------------------- -------------------------------------------------------- -Time: 0.488s Load: 0.008s, Pack+Encode: 0.200s, Decode+Unpack: 0.281s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 124, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 124, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7568 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample38-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample38-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample388-layer4-item1.zst (53/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample388-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,644B, BPFP=0.2430 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 16,728B, BPFP=1.5375 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,904B, BPFP=0.8184 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,844B, BPFP=1.6401 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,376B, BPFP=1.0456 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 18,724B, BPFP=1.7210 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,908B, BPFP=1.0945 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 18,056B, BPFP=1.6596 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,468B, BPFP=0.7783 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,064B, BPFP=1.7522 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 35,228B, BPFP=0.8095 -⌛️ [2/4] FRONTEND: Frontend time: 0.202s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02690007 27.36035156 - layer.0.v_cache 0.00000028 0.00032372 - layer.1.k_cache 0.00337091 1.74300932 - layer.1.v_cache 0.00000080 0.00106004 - layer.2.k_cache 0.00114287 0.77989430 - layer.2.v_cache 0.00000105 0.00150887 - layer.3.k_cache 0.00139664 0.87109626 - layer.3.v_cache 0.00000196 0.00246411 - layer.4.k_cache 0.00332559 1.90638051 - layer.4.v_cache 0.00000294 0.00420682 - layer.4.output 0.00025832 0.14783289 - ------------------------------------------------------------------------------------- - TOTAL 0.00265546 2.37583051 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 168944 -BPFP 1.1091 bits/point -EBPFP 2.2183 equivalent bits/point -MSE 2.375831 ----------------------- -------------------------------------------------------- -Time: 0.485s Load: 0.006s, Pack+Encode: 0.202s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.3758 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample388-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample388-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample390-layer4-item1.zst (54/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample390-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 74, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 74, 128) -Output shape: (1, 74, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.0.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.1.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.1.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.2.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.2.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.3.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.3.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.4.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.4.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.4.output: torch.Size([1, 74, 4096]) -> torch.Size([1, 1, 74, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,460B, BPFP=0.2597 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 15,656B, BPFP=1.6529 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,080B, BPFP=0.8530 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,064B, BPFP=1.8015 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,236B, BPFP=1.1862 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 17,900B, BPFP=1.8898 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,684B, BPFP=1.1280 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 16,388B, BPFP=1.7302 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,164B, BPFP=0.7563 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 17,208B, BPFP=1.8167 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 36,196B, BPFP=0.9553 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 74, 128]) - layer.0.v_cache: torch.Size([1, 8, 74, 128]) - layer.1.k_cache: torch.Size([1, 8, 74, 128]) - layer.1.v_cache: torch.Size([1, 8, 74, 128]) - layer.2.k_cache: torch.Size([1, 8, 74, 128]) - layer.2.v_cache: torch.Size([1, 8, 74, 128]) - layer.3.k_cache: torch.Size([1, 8, 74, 128]) - layer.3.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.k_cache: torch.Size([1, 8, 74, 128]) - layer.4.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.output: torch.Size([1, 74, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.274s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 74, 128]) - layer.0.v_cache: torch.Size([1, 8, 74, 128]) - layer.1.k_cache: torch.Size([1, 8, 74, 128]) - layer.1.v_cache: torch.Size([1, 8, 74, 128]) - layer.2.k_cache: torch.Size([1, 8, 74, 128]) - layer.2.v_cache: torch.Size([1, 8, 74, 128]) - layer.3.k_cache: torch.Size([1, 8, 74, 128]) - layer.3.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.k_cache: torch.Size([1, 8, 74, 128]) - layer.4.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.output: torch.Size([1, 74, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02896725 32.88911067 - layer.0.v_cache 0.00000027 0.00032505 - layer.1.k_cache 0.00342263 1.84460037 - layer.1.v_cache 0.00000079 0.00121221 - layer.2.k_cache 0.00118243 0.78451858 - layer.2.v_cache 0.00000111 0.00164959 - layer.3.k_cache 0.00141981 0.91294510 - layer.3.v_cache 0.00000233 0.00264796 - layer.4.k_cache 0.00326747 1.84747232 - layer.4.v_cache 0.00000299 0.00434112 - layer.4.output 0.00023783 0.14808129 - ------------------------------------------------------------------------------------- - TOTAL 0.00280131 2.77722487 - (elements=1,060,864) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1060864 -Total Bytes 160036 -BPFP 1.2068 bits/point -EBPFP 2.4137 equivalent bits/point -MSE 2.777225 ----------------------- -------------------------------------------------------- -Time: 0.478s Load: 0.005s, Pack+Encode: 0.200s, Decode+Unpack: 0.274s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 74, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.7772 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample390-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample390-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample40-layer4-item1.zst (55/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample40-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 126, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 126, 128) -Output shape: (1, 126, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.0.v_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.1.k_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.1.v_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.2.k_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.2.v_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.3.k_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.3.v_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.4.k_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.4.v_cache: torch.Size([1, 8, 126, 128]) -> torch.Size([1, 1, 126, 1024]) - layer.4.output: torch.Size([1, 126, 4096]) -> torch.Size([1, 1, 126, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,576B, BPFP=0.2217 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,996B, BPFP=1.3018 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,584B, BPFP=0.6562 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,808B, BPFP=1.4142 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,184B, BPFP=0.8175 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,460B, BPFP=1.3926 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,932B, BPFP=0.8638 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,716B, BPFP=1.3465 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,296B, BPFP=0.6384 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,076B, BPFP=1.4308 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 50,680B, BPFP=0.7856 -⌛️ [2/4] FRONTEND: Frontend time: 0.201s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 126, 128]) - layer.0.v_cache: torch.Size([1, 8, 126, 128]) - layer.1.k_cache: torch.Size([1, 8, 126, 128]) - layer.1.v_cache: torch.Size([1, 8, 126, 128]) - layer.2.k_cache: torch.Size([1, 8, 126, 128]) - layer.2.v_cache: torch.Size([1, 8, 126, 128]) - layer.3.k_cache: torch.Size([1, 8, 126, 128]) - layer.3.v_cache: torch.Size([1, 8, 126, 128]) - layer.4.k_cache: torch.Size([1, 8, 126, 128]) - layer.4.v_cache: torch.Size([1, 8, 126, 128]) - layer.4.output: torch.Size([1, 126, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.279s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 126, 128]) - layer.0.v_cache: torch.Size([1, 8, 126, 128]) - layer.1.k_cache: torch.Size([1, 8, 126, 128]) - layer.1.v_cache: torch.Size([1, 8, 126, 128]) - layer.2.k_cache: torch.Size([1, 8, 126, 128]) - layer.2.v_cache: torch.Size([1, 8, 126, 128]) - layer.3.k_cache: torch.Size([1, 8, 126, 128]) - layer.3.v_cache: torch.Size([1, 8, 126, 128]) - layer.4.k_cache: torch.Size([1, 8, 126, 128]) - layer.4.v_cache: torch.Size([1, 8, 126, 128]) - layer.4.output: torch.Size([1, 126, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02675629 19.05170743 - layer.0.v_cache 0.00000026 0.00029259 - layer.1.k_cache 0.00313964 1.38604179 - layer.1.v_cache 0.00000077 0.00106209 - layer.2.k_cache 0.00115655 0.71363382 - layer.2.v_cache 0.00000107 0.00151647 - layer.3.k_cache 0.00144242 0.82659379 - layer.3.v_cache 0.00000205 0.00250158 - layer.4.k_cache 0.00340494 1.65505352 - layer.4.v_cache 0.00000295 0.00415869 - layer.4.output 0.00023638 0.11419608 - ------------------------------------------------------------------------------------- - TOTAL 0.00263232 1.72138186 - (elements=1,806,336) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1806336 -Total Bytes 213308 -BPFP 0.9447 bits/point -EBPFP 1.8894 equivalent bits/point -MSE 1.721382 ----------------------- -------------------------------------------------------- -Time: 0.487s Load: 0.007s, Pack+Encode: 0.201s, Decode+Unpack: 0.279s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 126, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 126, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7214 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample40-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample40-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample412-layer4-item1.zst (56/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample412-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 76, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 76, 128) -Output shape: (1, 76, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.0.v_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.1.k_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.1.v_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.2.k_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.2.v_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.3.k_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.3.v_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.4.k_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.4.v_cache: torch.Size([1, 8, 76, 128]) -> torch.Size([1, 1, 76, 1024]) - layer.4.output: torch.Size([1, 76, 4096]) -> torch.Size([1, 1, 76, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,504B, BPFP=0.2574 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 15,364B, BPFP=1.5794 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,080B, BPFP=0.8306 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 16,056B, BPFP=1.6505 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,712B, BPFP=1.1012 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,096B, BPFP=1.6546 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,996B, BPFP=1.1303 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 16,396B, BPFP=1.6854 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,124B, BPFP=0.7323 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 16,468B, BPFP=1.6928 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 35,044B, BPFP=0.9006 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 76, 128]) - layer.0.v_cache: torch.Size([1, 8, 76, 128]) - layer.1.k_cache: torch.Size([1, 8, 76, 128]) - layer.1.v_cache: torch.Size([1, 8, 76, 128]) - layer.2.k_cache: torch.Size([1, 8, 76, 128]) - layer.2.v_cache: torch.Size([1, 8, 76, 128]) - layer.3.k_cache: torch.Size([1, 8, 76, 128]) - layer.3.v_cache: torch.Size([1, 8, 76, 128]) - layer.4.k_cache: torch.Size([1, 8, 76, 128]) - layer.4.v_cache: torch.Size([1, 8, 76, 128]) - layer.4.output: torch.Size([1, 76, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 76, 128]) - layer.0.v_cache: torch.Size([1, 8, 76, 128]) - layer.1.k_cache: torch.Size([1, 8, 76, 128]) - layer.1.v_cache: torch.Size([1, 8, 76, 128]) - layer.2.k_cache: torch.Size([1, 8, 76, 128]) - layer.2.v_cache: torch.Size([1, 8, 76, 128]) - layer.3.k_cache: torch.Size([1, 8, 76, 128]) - layer.3.v_cache: torch.Size([1, 8, 76, 128]) - layer.4.k_cache: torch.Size([1, 8, 76, 128]) - layer.4.v_cache: torch.Size([1, 8, 76, 128]) - layer.4.output: torch.Size([1, 76, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02706527 31.83264481 - layer.0.v_cache 0.00000027 0.00030737 - layer.1.k_cache 0.00360181 1.85852292 - layer.1.v_cache 0.00000073 0.00103544 - layer.2.k_cache 0.00113352 0.81087102 - layer.2.v_cache 0.00000108 0.00158674 - layer.3.k_cache 0.00139474 0.91066220 - layer.3.v_cache 0.00000224 0.00259610 - layer.4.k_cache 0.00318908 1.92290176 - layer.4.v_cache 0.00000296 0.00426294 - layer.4.output 0.00022408 0.16402029 - ------------------------------------------------------------------------------------- - TOTAL 0.00266343 2.71439089 - (elements=1,089,536) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1089536 -Total Bytes 154840 -BPFP 1.1369 bits/point -EBPFP 2.2738 equivalent bits/point -MSE 2.714391 ----------------------- -------------------------------------------------------- -Time: 0.482s Load: 0.006s, Pack+Encode: 0.200s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 76, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 76, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.7144 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample412-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample412-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample414-layer4-item1.zst (57/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample414-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,716B, BPFP=0.2496 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 17,776B, BPFP=1.6338 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,880B, BPFP=0.9081 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,632B, BPFP=1.7125 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,240B, BPFP=1.1250 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 18,852B, BPFP=1.7327 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 12,132B, BPFP=1.1151 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 18,532B, BPFP=1.7033 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,788B, BPFP=0.8077 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 18,608B, BPFP=1.7103 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 37,288B, BPFP=0.8568 -⌛️ [2/4] FRONTEND: Frontend time: 0.197s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02461138 26.53386949 - layer.0.v_cache 0.00000028 0.00030755 - layer.1.k_cache 0.00340719 1.65702084 - layer.1.v_cache 0.00000073 0.00099673 - layer.2.k_cache 0.00115796 0.76746018 - layer.2.v_cache 0.00000109 0.00147199 - layer.3.k_cache 0.00135314 0.86443535 - layer.3.v_cache 0.00000196 0.00239261 - layer.4.k_cache 0.00328730 1.70225004 - layer.4.v_cache 0.00000286 0.00404241 - layer.4.output 0.00022769 0.13889861 - ------------------------------------------------------------------------------------- - TOTAL 0.00248105 2.29213155 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 175444 -BPFP 1.1518 bits/point -EBPFP 2.3036 equivalent bits/point -MSE 2.292132 ----------------------- -------------------------------------------------------- -Time: 0.480s Load: 0.006s, Pack+Encode: 0.197s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2921 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample414-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample414-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample42-layer4-item1.zst (58/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample42-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 134, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 134, 128) -Output shape: (1, 134, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.0.v_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.1.k_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.1.v_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.2.k_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.2.v_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.3.k_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.3.v_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.4.k_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.4.v_cache: torch.Size([1, 8, 134, 128]) -> torch.Size([1, 1, 134, 1024]) - layer.4.output: torch.Size([1, 134, 4096]) -> torch.Size([1, 1, 134, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,908B, BPFP=0.2861 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 27,276B, BPFP=1.5903 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 15,180B, BPFP=0.8850 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 29,068B, BPFP=1.6947 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 17,748B, BPFP=1.0347 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 29,212B, BPFP=1.7031 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,328B, BPFP=1.1269 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 29,092B, BPFP=1.6961 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 15,336B, BPFP=0.8941 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 29,884B, BPFP=1.7423 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 67,004B, BPFP=0.9766 -⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 134, 128]) - layer.0.v_cache: torch.Size([1, 8, 134, 128]) - layer.1.k_cache: torch.Size([1, 8, 134, 128]) - layer.1.v_cache: torch.Size([1, 8, 134, 128]) - layer.2.k_cache: torch.Size([1, 8, 134, 128]) - layer.2.v_cache: torch.Size([1, 8, 134, 128]) - layer.3.k_cache: torch.Size([1, 8, 134, 128]) - layer.3.v_cache: torch.Size([1, 8, 134, 128]) - layer.4.k_cache: torch.Size([1, 8, 134, 128]) - layer.4.v_cache: torch.Size([1, 8, 134, 128]) - layer.4.output: torch.Size([1, 134, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.372s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 134, 128]) - layer.0.v_cache: torch.Size([1, 8, 134, 128]) - layer.1.k_cache: torch.Size([1, 8, 134, 128]) - layer.1.v_cache: torch.Size([1, 8, 134, 128]) - layer.2.k_cache: torch.Size([1, 8, 134, 128]) - layer.2.v_cache: torch.Size([1, 8, 134, 128]) - layer.3.k_cache: torch.Size([1, 8, 134, 128]) - layer.3.v_cache: torch.Size([1, 8, 134, 128]) - layer.4.k_cache: torch.Size([1, 8, 134, 128]) - layer.4.v_cache: torch.Size([1, 8, 134, 128]) - layer.4.output: torch.Size([1, 134, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02623714 22.88941705 - layer.0.v_cache 0.00000029 0.00030338 - layer.1.k_cache 0.00318200 1.62077890 - layer.1.v_cache 0.00000078 0.00099056 - layer.2.k_cache 0.00115661 0.72619424 - layer.2.v_cache 0.00000102 0.00136474 - layer.3.k_cache 0.00142778 0.83116127 - layer.3.v_cache 0.00000209 0.00242744 - layer.4.k_cache 0.00344207 1.66925709 - layer.4.v_cache 0.00000292 0.00394332 - layer.4.output 0.00019652 0.11952558 - ------------------------------------------------------------------------------------- - TOTAL 0.00258848 2.01599574 - (elements=1,921,024) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1921024 -Total Bytes 284036 -BPFP 1.1829 bits/point -EBPFP 2.3657 equivalent bits/point -MSE 2.015996 ----------------------- -------------------------------------------------------- -Time: 0.625s Load: 0.008s, Pack+Encode: 0.245s, Decode+Unpack: 0.372s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 134, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 134, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0160 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample42-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample42-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample44-layer4-item1.zst (59/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample44-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 132, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 132, 128) -Output shape: (1, 132, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.0.v_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.1.k_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.1.v_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.2.k_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.2.v_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.3.k_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.3.v_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.4.k_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.4.v_cache: torch.Size([1, 8, 132, 128]) -> torch.Size([1, 1, 132, 1024]) - layer.4.output: torch.Size([1, 132, 4096]) -> torch.Size([1, 1, 132, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,672B, BPFP=0.2765 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 27,264B, BPFP=1.6136 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 15,764B, BPFP=0.9330 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 29,216B, BPFP=1.7292 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,096B, BPFP=1.0710 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 29,720B, BPFP=1.7590 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,904B, BPFP=1.1188 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 29,300B, BPFP=1.7341 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,028B, BPFP=1.0078 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 29,812B, BPFP=1.7644 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 88,924B, BPFP=1.3158 -⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 132, 128]) - layer.0.v_cache: torch.Size([1, 8, 132, 128]) - layer.1.k_cache: torch.Size([1, 8, 132, 128]) - layer.1.v_cache: torch.Size([1, 8, 132, 128]) - layer.2.k_cache: torch.Size([1, 8, 132, 128]) - layer.2.v_cache: torch.Size([1, 8, 132, 128]) - layer.3.k_cache: torch.Size([1, 8, 132, 128]) - layer.3.v_cache: torch.Size([1, 8, 132, 128]) - layer.4.k_cache: torch.Size([1, 8, 132, 128]) - layer.4.v_cache: torch.Size([1, 8, 132, 128]) - layer.4.output: torch.Size([1, 132, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.376s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 132, 128]) - layer.0.v_cache: torch.Size([1, 8, 132, 128]) - layer.1.k_cache: torch.Size([1, 8, 132, 128]) - layer.1.v_cache: torch.Size([1, 8, 132, 128]) - layer.2.k_cache: torch.Size([1, 8, 132, 128]) - layer.2.v_cache: torch.Size([1, 8, 132, 128]) - layer.3.k_cache: torch.Size([1, 8, 132, 128]) - layer.3.v_cache: torch.Size([1, 8, 132, 128]) - layer.4.k_cache: torch.Size([1, 8, 132, 128]) - layer.4.v_cache: torch.Size([1, 8, 132, 128]) - layer.4.output: torch.Size([1, 132, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02542288 22.62650183 - layer.0.v_cache 0.00000027 0.00030292 - layer.1.k_cache 0.00323066 1.53819622 - layer.1.v_cache 0.00000090 0.00109242 - layer.2.k_cache 0.00115984 0.72998868 - layer.2.v_cache 0.00000123 0.00162902 - layer.3.k_cache 0.00138390 0.82726305 - layer.3.v_cache 0.00000215 0.00269264 - layer.4.k_cache 0.00364779 1.65702057 - layer.4.v_cache 0.00000297 0.00413128 - layer.4.output 0.00020533 0.12268498 - ------------------------------------------------------------------------------------- - TOTAL 0.00254814 1.99139704 - (elements=1,892,352) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1892352 -Total Bytes 308700 -BPFP 1.3050 bits/point -EBPFP 2.6101 equivalent bits/point -MSE 1.991397 ----------------------- -------------------------------------------------------- -Time: 0.629s Load: 0.007s, Pack+Encode: 0.246s, Decode+Unpack: 0.376s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 132, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 132, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9914 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample44-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample44-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample454-layer4-item1.zst (60/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample454-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 77, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 77, 128) -Output shape: (1, 77, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.0.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.1.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.1.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.2.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.2.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.3.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.3.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.4.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.4.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.4.output: torch.Size([1, 77, 4096]) -> torch.Size([1, 1, 77, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,556B, BPFP=0.2593 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 15,876B, BPFP=1.6108 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,576B, BPFP=0.8701 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,344B, BPFP=1.7597 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,376B, BPFP=1.1542 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 17,920B, BPFP=1.8182 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,436B, BPFP=1.1603 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,412B, BPFP=1.7666 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,664B, BPFP=0.7776 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 17,636B, BPFP=1.7894 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,828B, BPFP=1.0356 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 77, 128]) - layer.0.v_cache: torch.Size([1, 8, 77, 128]) - layer.1.k_cache: torch.Size([1, 8, 77, 128]) - layer.1.v_cache: torch.Size([1, 8, 77, 128]) - layer.2.k_cache: torch.Size([1, 8, 77, 128]) - layer.2.v_cache: torch.Size([1, 8, 77, 128]) - layer.3.k_cache: torch.Size([1, 8, 77, 128]) - layer.3.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.k_cache: torch.Size([1, 8, 77, 128]) - layer.4.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.output: torch.Size([1, 77, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 77, 128]) - layer.0.v_cache: torch.Size([1, 8, 77, 128]) - layer.1.k_cache: torch.Size([1, 8, 77, 128]) - layer.1.v_cache: torch.Size([1, 8, 77, 128]) - layer.2.k_cache: torch.Size([1, 8, 77, 128]) - layer.2.v_cache: torch.Size([1, 8, 77, 128]) - layer.3.k_cache: torch.Size([1, 8, 77, 128]) - layer.3.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.k_cache: torch.Size([1, 8, 77, 128]) - layer.4.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.output: torch.Size([1, 77, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02625101 29.79534357 - layer.0.v_cache 0.00000028 0.00032970 - layer.1.k_cache 0.00355533 1.80405297 - layer.1.v_cache 0.00000093 0.00117608 - layer.2.k_cache 0.00115642 0.77199782 - layer.2.v_cache 0.00000119 0.00169269 - layer.3.k_cache 0.00136722 0.87838814 - layer.3.v_cache 0.00000230 0.00282068 - layer.4.k_cache 0.00337541 1.74096026 - layer.4.v_cache 0.00000304 0.00460909 - layer.4.output 0.00023573 0.13494505 - ------------------------------------------------------------------------------------- - TOTAL 0.00261829 2.53865366 - (elements=1,103,872) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1103872 -Total Bytes 168624 -BPFP 1.2221 bits/point -EBPFP 2.4441 equivalent bits/point -MSE 2.538654 ----------------------- -------------------------------------------------------- -Time: 0.481s Load: 0.006s, Pack+Encode: 0.199s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 77, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5387 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample454-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample454-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample464-layer4-item1.zst (61/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample464-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,848B, BPFP=0.2367 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,068B, BPFP=1.6679 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,440B, BPFP=0.8677 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 20,632B, BPFP=1.7148 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,348B, BPFP=1.0263 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 20,864B, BPFP=1.7340 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,184B, BPFP=1.0957 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 20,752B, BPFP=1.7247 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,180B, BPFP=0.8461 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 20,228B, BPFP=1.6812 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 55,612B, BPFP=1.1555 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.275s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02258782 25.28207073 - layer.0.v_cache 0.00000029 0.00029849 - layer.1.k_cache 0.00339276 1.68054491 - layer.1.v_cache 0.00000074 0.00093593 - layer.2.k_cache 0.00116230 0.72565420 - layer.2.v_cache 0.00000115 0.00142025 - layer.3.k_cache 0.00135573 0.80563858 - layer.3.v_cache 0.00000225 0.00253363 - layer.4.k_cache 0.00337629 1.67671009 - layer.4.v_cache 0.00000290 0.00383583 - layer.4.output 0.00017695 0.12421046 - ------------------------------------------------------------------------------------- - TOTAL 0.00232786 2.19117746 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 207156 -BPFP 1.2298 bits/point -EBPFP 2.4596 equivalent bits/point -MSE 2.191177 ----------------------- -------------------------------------------------------- -Time: 0.479s Load: 0.005s, Pack+Encode: 0.199s, Decode+Unpack: 0.275s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1912 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample464-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample464-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample478-layer4-item1.zst (62/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample478-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 77, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 77, 128) -Output shape: (1, 77, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.0.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.1.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.1.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.2.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.2.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.3.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.3.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.4.k_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.4.v_cache: torch.Size([1, 8, 77, 128]) -> torch.Size([1, 1, 77, 1024]) - layer.4.output: torch.Size([1, 77, 4096]) -> torch.Size([1, 1, 77, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,560B, BPFP=0.2597 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 16,296B, BPFP=1.6534 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,356B, BPFP=0.8478 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,012B, BPFP=1.7261 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,944B, BPFP=1.1104 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 17,468B, BPFP=1.7723 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,100B, BPFP=1.1262 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,100B, BPFP=1.7350 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,588B, BPFP=0.7699 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 17,068B, BPFP=1.7317 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 34,768B, BPFP=0.8819 -⌛️ [2/4] FRONTEND: Frontend time: 0.203s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 77, 128]) - layer.0.v_cache: torch.Size([1, 8, 77, 128]) - layer.1.k_cache: torch.Size([1, 8, 77, 128]) - layer.1.v_cache: torch.Size([1, 8, 77, 128]) - layer.2.k_cache: torch.Size([1, 8, 77, 128]) - layer.2.v_cache: torch.Size([1, 8, 77, 128]) - layer.3.k_cache: torch.Size([1, 8, 77, 128]) - layer.3.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.k_cache: torch.Size([1, 8, 77, 128]) - layer.4.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.output: torch.Size([1, 77, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 77, 128]) - layer.0.v_cache: torch.Size([1, 8, 77, 128]) - layer.1.k_cache: torch.Size([1, 8, 77, 128]) - layer.1.v_cache: torch.Size([1, 8, 77, 128]) - layer.2.k_cache: torch.Size([1, 8, 77, 128]) - layer.2.v_cache: torch.Size([1, 8, 77, 128]) - layer.3.k_cache: torch.Size([1, 8, 77, 128]) - layer.3.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.k_cache: torch.Size([1, 8, 77, 128]) - layer.4.v_cache: torch.Size([1, 8, 77, 128]) - layer.4.output: torch.Size([1, 77, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02707502 28.71634030 - layer.0.v_cache 0.00000027 0.00032547 - layer.1.k_cache 0.00339772 1.81974892 - layer.1.v_cache 0.00000076 0.00114445 - layer.2.k_cache 0.00114556 0.80254017 - layer.2.v_cache 0.00000112 0.00170318 - layer.3.k_cache 0.00146787 0.90835720 - layer.3.v_cache 0.00000204 0.00274815 - layer.4.k_cache 0.00318992 1.87572588 - layer.4.v_cache 0.00000300 0.00460624 - layer.4.output 0.00024419 0.16627523 - ------------------------------------------------------------------------------------- - TOTAL 0.00266143 2.48559578 - (elements=1,103,872) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1103872 -Total Bytes 160260 -BPFP 1.1614 bits/point -EBPFP 2.3229 equivalent bits/point -MSE 2.485596 ----------------------- -------------------------------------------------------- -Time: 0.485s Load: 0.006s, Pack+Encode: 0.203s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 77, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 77, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4856 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample478-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample478-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample48-layer4-item1.zst (63/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample48-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 112, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 112, 128) -Output shape: (1, 112, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.0.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.1.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.1.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.2.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.2.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.3.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.3.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.4.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.4.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.4.output: torch.Size([1, 112, 4096]) -> torch.Size([1, 1, 112, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,316B, BPFP=0.2313 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,668B, BPFP=1.4417 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,696B, BPFP=0.6763 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,224B, BPFP=1.5502 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,548B, BPFP=0.8753 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,144B, BPFP=1.5446 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,408B, BPFP=0.9353 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,380B, BPFP=1.4914 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,820B, BPFP=0.7547 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,816B, BPFP=1.5915 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 48,756B, BPFP=0.8502 -⌛️ [2/4] FRONTEND: Frontend time: 0.198s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 112, 128]) - layer.0.v_cache: torch.Size([1, 8, 112, 128]) - layer.1.k_cache: torch.Size([1, 8, 112, 128]) - layer.1.v_cache: torch.Size([1, 8, 112, 128]) - layer.2.k_cache: torch.Size([1, 8, 112, 128]) - layer.2.v_cache: torch.Size([1, 8, 112, 128]) - layer.3.k_cache: torch.Size([1, 8, 112, 128]) - layer.3.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.k_cache: torch.Size([1, 8, 112, 128]) - layer.4.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.output: torch.Size([1, 112, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 112, 128]) - layer.0.v_cache: torch.Size([1, 8, 112, 128]) - layer.1.k_cache: torch.Size([1, 8, 112, 128]) - layer.1.v_cache: torch.Size([1, 8, 112, 128]) - layer.2.k_cache: torch.Size([1, 8, 112, 128]) - layer.2.v_cache: torch.Size([1, 8, 112, 128]) - layer.3.k_cache: torch.Size([1, 8, 112, 128]) - layer.3.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.k_cache: torch.Size([1, 8, 112, 128]) - layer.4.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.output: torch.Size([1, 112, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02677278 22.05531965 - layer.0.v_cache 0.00000026 0.00030173 - layer.1.k_cache 0.00327242 1.54847963 - layer.1.v_cache 0.00000072 0.00104898 - layer.2.k_cache 0.00117792 0.75829533 - layer.2.v_cache 0.00000108 0.00154956 - layer.3.k_cache 0.00137547 0.84576750 - layer.3.v_cache 0.00000205 0.00256288 - layer.4.k_cache 0.00331726 1.79367570 - layer.4.v_cache 0.00000297 0.00417305 - layer.4.output 0.00024922 0.12549341 - ------------------------------------------------------------------------------------- - TOTAL 0.00263713 1.96522483 - (elements=1,605,632) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1605632 -Total Bytes 207776 -BPFP 1.0352 bits/point -EBPFP 2.0705 equivalent bits/point -MSE 1.965225 ----------------------- -------------------------------------------------------- -Time: 0.482s Load: 0.007s, Pack+Encode: 0.198s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 112, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9652 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample48-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample48-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample485-layer4-item1.zst (64/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample485-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 78, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 78, 128) -Output shape: (1, 78, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.0.v_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.1.k_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.1.v_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.2.k_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.2.v_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.3.k_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.3.v_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.4.k_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.4.v_cache: torch.Size([1, 8, 78, 128]) -> torch.Size([1, 1, 78, 1024]) - layer.4.output: torch.Size([1, 78, 4096]) -> torch.Size([1, 1, 78, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,596B, BPFP=0.2600 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 15,128B, BPFP=1.5152 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,584B, BPFP=0.8598 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 16,712B, BPFP=1.6739 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,336B, BPFP=1.2356 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 17,752B, BPFP=1.7780 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,556B, BPFP=1.1575 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,316B, BPFP=1.7344 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,672B, BPFP=0.7684 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 18,052B, BPFP=1.8081 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,372B, BPFP=0.9608 -⌛️ [2/4] FRONTEND: Frontend time: 0.201s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 78, 128]) - layer.0.v_cache: torch.Size([1, 8, 78, 128]) - layer.1.k_cache: torch.Size([1, 8, 78, 128]) - layer.1.v_cache: torch.Size([1, 8, 78, 128]) - layer.2.k_cache: torch.Size([1, 8, 78, 128]) - layer.2.v_cache: torch.Size([1, 8, 78, 128]) - layer.3.k_cache: torch.Size([1, 8, 78, 128]) - layer.3.v_cache: torch.Size([1, 8, 78, 128]) - layer.4.k_cache: torch.Size([1, 8, 78, 128]) - layer.4.v_cache: torch.Size([1, 8, 78, 128]) - layer.4.output: torch.Size([1, 78, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 78, 128]) - layer.0.v_cache: torch.Size([1, 8, 78, 128]) - layer.1.k_cache: torch.Size([1, 8, 78, 128]) - layer.1.v_cache: torch.Size([1, 8, 78, 128]) - layer.2.k_cache: torch.Size([1, 8, 78, 128]) - layer.2.v_cache: torch.Size([1, 8, 78, 128]) - layer.3.k_cache: torch.Size([1, 8, 78, 128]) - layer.3.v_cache: torch.Size([1, 8, 78, 128]) - layer.4.k_cache: torch.Size([1, 8, 78, 128]) - layer.4.v_cache: torch.Size([1, 8, 78, 128]) - layer.4.output: torch.Size([1, 78, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02543071 26.75562149 - layer.0.v_cache 0.00000030 0.00031471 - layer.1.k_cache 0.00350441 1.85771570 - layer.1.v_cache 0.00000082 0.00115692 - layer.2.k_cache 0.00114279 0.77866975 - layer.2.v_cache 0.00000132 0.00177158 - layer.3.k_cache 0.00135780 0.88222572 - layer.3.v_cache 0.00000231 0.00291236 - layer.4.k_cache 0.00328533 1.82815004 - layer.4.v_cache 0.00000329 0.00487361 - layer.4.output 0.00022272 0.13478804 - ------------------------------------------------------------------------------------- - TOTAL 0.00254428 2.33232600 - (elements=1,118,208) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1118208 -Total Bytes 166076 -BPFP 1.1882 bits/point -EBPFP 2.3763 equivalent bits/point -MSE 2.332326 ----------------------- -------------------------------------------------------- -Time: 0.484s Load: 0.005s, Pack+Encode: 0.201s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 78, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 78, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.3323 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample485-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample485-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample487-layer4-item1.zst (65/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample487-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 79, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 79, 128) -Output shape: (1, 79, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.0.v_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.1.k_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.1.v_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.2.k_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.2.v_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.3.k_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.3.v_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.4.k_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.4.v_cache: torch.Size([1, 8, 79, 128]) -> torch.Size([1, 1, 79, 1024]) - layer.4.output: torch.Size([1, 79, 4096]) -> torch.Size([1, 1, 79, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,608B, BPFP=0.2579 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 16,532B, BPFP=1.6349 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,724B, BPFP=0.8627 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 18,164B, BPFP=1.7963 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,608B, BPFP=1.1479 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 18,880B, BPFP=1.8671 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,836B, BPFP=1.1705 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 18,992B, BPFP=1.8782 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,948B, BPFP=0.7860 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 19,320B, BPFP=1.9106 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 32,932B, BPFP=0.8142 -⌛️ [2/4] FRONTEND: Frontend time: 0.198s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 79, 128]) - layer.0.v_cache: torch.Size([1, 8, 79, 128]) - layer.1.k_cache: torch.Size([1, 8, 79, 128]) - layer.1.v_cache: torch.Size([1, 8, 79, 128]) - layer.2.k_cache: torch.Size([1, 8, 79, 128]) - layer.2.v_cache: torch.Size([1, 8, 79, 128]) - layer.3.k_cache: torch.Size([1, 8, 79, 128]) - layer.3.v_cache: torch.Size([1, 8, 79, 128]) - layer.4.k_cache: torch.Size([1, 8, 79, 128]) - layer.4.v_cache: torch.Size([1, 8, 79, 128]) - layer.4.output: torch.Size([1, 79, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.277s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 79, 128]) - layer.0.v_cache: torch.Size([1, 8, 79, 128]) - layer.1.k_cache: torch.Size([1, 8, 79, 128]) - layer.1.v_cache: torch.Size([1, 8, 79, 128]) - layer.2.k_cache: torch.Size([1, 8, 79, 128]) - layer.2.v_cache: torch.Size([1, 8, 79, 128]) - layer.3.k_cache: torch.Size([1, 8, 79, 128]) - layer.3.v_cache: torch.Size([1, 8, 79, 128]) - layer.4.k_cache: torch.Size([1, 8, 79, 128]) - layer.4.v_cache: torch.Size([1, 8, 79, 128]) - layer.4.output: torch.Size([1, 79, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02536859 28.83384531 - layer.0.v_cache 0.00000029 0.00031844 - layer.1.k_cache 0.00331158 1.81490278 - layer.1.v_cache 0.00000077 0.00111079 - layer.2.k_cache 0.00112976 0.75026384 - layer.2.v_cache 0.00000111 0.00158401 - layer.3.k_cache 0.00133964 0.86268790 - layer.3.v_cache 0.00000223 0.00267613 - layer.4.k_cache 0.00332594 1.77556523 - layer.4.v_cache 0.00000299 0.00416082 - layer.4.output 0.00020400 0.13063475 - ------------------------------------------------------------------------------------- - TOTAL 0.00252135 2.46926102 - (elements=1,132,544) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1132544 -Total Bytes 167544 -BPFP 1.1835 bits/point -EBPFP 2.3670 equivalent bits/point -MSE 2.469261 ----------------------- -------------------------------------------------------- -Time: 0.481s Load: 0.005s, Pack+Encode: 0.198s, Decode+Unpack: 0.277s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 79, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 79, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4693 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample487-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample487-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample495-layer4-item1.zst (66/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample495-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 81, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 81, 128) -Output shape: (1, 81, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.0.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.1.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.1.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.2.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.2.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.3.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.3.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.4.k_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.4.v_cache: torch.Size([1, 8, 81, 128]) -> torch.Size([1, 1, 81, 1024]) - layer.4.output: torch.Size([1, 81, 4096]) -> torch.Size([1, 1, 81, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,636B, BPFP=0.2542 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 16,196B, BPFP=1.5621 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,068B, BPFP=0.8746 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,476B, BPFP=1.6856 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,584B, BPFP=1.1173 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 18,552B, BPFP=1.7894 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,912B, BPFP=1.1489 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 18,256B, BPFP=1.7608 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,396B, BPFP=0.8098 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 18,164B, BPFP=1.7519 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,412B, BPFP=1.1191 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 81, 128]) - layer.0.v_cache: torch.Size([1, 8, 81, 128]) - layer.1.k_cache: torch.Size([1, 8, 81, 128]) - layer.1.v_cache: torch.Size([1, 8, 81, 128]) - layer.2.k_cache: torch.Size([1, 8, 81, 128]) - layer.2.v_cache: torch.Size([1, 8, 81, 128]) - layer.3.k_cache: torch.Size([1, 8, 81, 128]) - layer.3.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.k_cache: torch.Size([1, 8, 81, 128]) - layer.4.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.output: torch.Size([1, 81, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 81, 128]) - layer.0.v_cache: torch.Size([1, 8, 81, 128]) - layer.1.k_cache: torch.Size([1, 8, 81, 128]) - layer.1.v_cache: torch.Size([1, 8, 81, 128]) - layer.2.k_cache: torch.Size([1, 8, 81, 128]) - layer.2.v_cache: torch.Size([1, 8, 81, 128]) - layer.3.k_cache: torch.Size([1, 8, 81, 128]) - layer.3.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.k_cache: torch.Size([1, 8, 81, 128]) - layer.4.v_cache: torch.Size([1, 8, 81, 128]) - layer.4.output: torch.Size([1, 81, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02658518 28.11517108 - layer.0.v_cache 0.00000029 0.00033245 - layer.1.k_cache 0.00348817 2.00764597 - layer.1.v_cache 0.00000087 0.00118823 - layer.2.k_cache 0.00114702 0.80974287 - layer.2.v_cache 0.00000128 0.00177944 - layer.3.k_cache 0.00135614 0.89564401 - layer.3.v_cache 0.00000252 0.00289205 - layer.4.k_cache 0.00328397 1.81980783 - layer.4.v_cache 0.00000321 0.00461545 - layer.4.output 0.00022006 0.14385634 - ------------------------------------------------------------------------------------- - TOTAL 0.00262492 2.44530320 - (elements=1,161,216) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1161216 -Total Bytes 178652 -BPFP 1.2308 bits/point -EBPFP 2.4616 equivalent bits/point -MSE 2.445303 ----------------------- -------------------------------------------------------- -Time: 0.495s Load: 0.007s, Pack+Encode: 0.199s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 81, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 81, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4453 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample495-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample495-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample50-layer4-item1.zst (67/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample50-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 117, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 117, 128) -Output shape: (1, 117, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.0.v_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.1.k_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.1.v_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.2.k_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.2.v_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.3.k_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.3.v_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.4.k_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.4.v_cache: torch.Size([1, 8, 117, 128]) -> torch.Size([1, 1, 117, 1024]) - layer.4.output: torch.Size([1, 117, 4096]) -> torch.Size([1, 1, 117, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,252B, BPFP=0.2171 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,860B, BPFP=1.3929 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,284B, BPFP=0.6867 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,300B, BPFP=1.5558 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,044B, BPFP=0.8710 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,980B, BPFP=1.5345 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,728B, BPFP=0.9167 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,072B, BPFP=1.4738 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,376B, BPFP=0.7596 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,308B, BPFP=1.5564 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 63,284B, BPFP=1.0564 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 117, 128]) - layer.0.v_cache: torch.Size([1, 8, 117, 128]) - layer.1.k_cache: torch.Size([1, 8, 117, 128]) - layer.1.v_cache: torch.Size([1, 8, 117, 128]) - layer.2.k_cache: torch.Size([1, 8, 117, 128]) - layer.2.v_cache: torch.Size([1, 8, 117, 128]) - layer.3.k_cache: torch.Size([1, 8, 117, 128]) - layer.3.v_cache: torch.Size([1, 8, 117, 128]) - layer.4.k_cache: torch.Size([1, 8, 117, 128]) - layer.4.v_cache: torch.Size([1, 8, 117, 128]) - layer.4.output: torch.Size([1, 117, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.290s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 117, 128]) - layer.0.v_cache: torch.Size([1, 8, 117, 128]) - layer.1.k_cache: torch.Size([1, 8, 117, 128]) - layer.1.v_cache: torch.Size([1, 8, 117, 128]) - layer.2.k_cache: torch.Size([1, 8, 117, 128]) - layer.2.v_cache: torch.Size([1, 8, 117, 128]) - layer.3.k_cache: torch.Size([1, 8, 117, 128]) - layer.3.v_cache: torch.Size([1, 8, 117, 128]) - layer.4.k_cache: torch.Size([1, 8, 117, 128]) - layer.4.v_cache: torch.Size([1, 8, 117, 128]) - layer.4.output: torch.Size([1, 117, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02657393 21.87872471 - layer.0.v_cache 0.00000026 0.00029695 - layer.1.k_cache 0.00326448 1.47496594 - layer.1.v_cache 0.00000079 0.00108250 - layer.2.k_cache 0.00117087 0.73288616 - layer.2.v_cache 0.00000122 0.00162230 - layer.3.k_cache 0.00141366 0.82635107 - layer.3.v_cache 0.00000241 0.00267005 - layer.4.k_cache 0.00339367 1.68755060 - layer.4.v_cache 0.00000311 0.00430854 - layer.4.output 0.00025967 0.13168595 - ------------------------------------------------------------------------------------- - TOTAL 0.00263308 1.93837162 - (elements=1,677,312) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1677312 -Total Bytes 227488 -BPFP 1.0850 bits/point -EBPFP 2.1700 equivalent bits/point -MSE 1.938372 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.007s, Pack+Encode: 0.206s, Decode+Unpack: 0.290s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 117, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 117, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9384 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample50-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample50-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample516-layer4-item1.zst (68/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample516-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 84, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 84, 128) -Output shape: (1, 84, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.0.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.1.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.1.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.2.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.2.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.3.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.3.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.4.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.4.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.4.output: torch.Size([1, 84, 4096]) -> torch.Size([1, 1, 84, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,640B, BPFP=0.2455 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 16,448B, BPFP=1.5298 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,420B, BPFP=0.8761 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,876B, BPFP=1.6626 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,488B, BPFP=1.0685 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 18,220B, BPFP=1.6946 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 11,952B, BPFP=1.1116 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,908B, BPFP=1.6656 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,332B, BPFP=0.7749 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 18,884B, BPFP=1.7563 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,720B, BPFP=0.9235 -⌛️ [2/4] FRONTEND: Frontend time: 0.205s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 84, 128]) - layer.0.v_cache: torch.Size([1, 8, 84, 128]) - layer.1.k_cache: torch.Size([1, 8, 84, 128]) - layer.1.v_cache: torch.Size([1, 8, 84, 128]) - layer.2.k_cache: torch.Size([1, 8, 84, 128]) - layer.2.v_cache: torch.Size([1, 8, 84, 128]) - layer.3.k_cache: torch.Size([1, 8, 84, 128]) - layer.3.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.k_cache: torch.Size([1, 8, 84, 128]) - layer.4.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.output: torch.Size([1, 84, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 84, 128]) - layer.0.v_cache: torch.Size([1, 8, 84, 128]) - layer.1.k_cache: torch.Size([1, 8, 84, 128]) - layer.1.v_cache: torch.Size([1, 8, 84, 128]) - layer.2.k_cache: torch.Size([1, 8, 84, 128]) - layer.2.v_cache: torch.Size([1, 8, 84, 128]) - layer.3.k_cache: torch.Size([1, 8, 84, 128]) - layer.3.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.k_cache: torch.Size([1, 8, 84, 128]) - layer.4.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.output: torch.Size([1, 84, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02585370 28.19101097 - layer.0.v_cache 0.00000028 0.00032230 - layer.1.k_cache 0.00342466 1.75519053 - layer.1.v_cache 0.00000080 0.00113560 - layer.2.k_cache 0.00116354 0.76309477 - layer.2.v_cache 0.00000112 0.00160593 - layer.3.k_cache 0.00133515 0.84630521 - layer.3.v_cache 0.00000214 0.00269214 - layer.4.k_cache 0.00358854 1.71544538 - layer.4.v_cache 0.00000313 0.00434895 - layer.4.output 0.00016584 0.12720586 - ------------------------------------------------------------------------------------- - TOTAL 0.00257403 2.41356966 - (elements=1,204,224) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1204224 -Total Bytes 172888 -BPFP 1.1485 bits/point -EBPFP 2.2971 equivalent bits/point -MSE 2.413570 ----------------------- -------------------------------------------------------- -Time: 0.499s Load: 0.006s, Pack+Encode: 0.205s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 84, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.4136 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample516-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample516-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample52-layer4-item1.zst (69/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample52-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 133, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 133, 128) -Output shape: (1, 133, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.0.v_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.1.k_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.1.v_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.2.k_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.2.v_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.3.k_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.3.v_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.4.k_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.4.v_cache: torch.Size([1, 8, 133, 128]) -> torch.Size([1, 1, 133, 1024]) - layer.4.output: torch.Size([1, 133, 4096]) -> torch.Size([1, 1, 133, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,780B, BPFP=0.2808 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 27,120B, BPFP=1.5930 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 15,476B, BPFP=0.9091 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 29,380B, BPFP=1.7258 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,312B, BPFP=1.0757 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 29,360B, BPFP=1.7246 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,936B, BPFP=1.1123 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 29,048B, BPFP=1.7063 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 15,736B, BPFP=0.9243 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 29,788B, BPFP=1.7498 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,448B, BPFP=1.1667 -⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 133, 128]) - layer.0.v_cache: torch.Size([1, 8, 133, 128]) - layer.1.k_cache: torch.Size([1, 8, 133, 128]) - layer.1.v_cache: torch.Size([1, 8, 133, 128]) - layer.2.k_cache: torch.Size([1, 8, 133, 128]) - layer.2.v_cache: torch.Size([1, 8, 133, 128]) - layer.3.k_cache: torch.Size([1, 8, 133, 128]) - layer.3.v_cache: torch.Size([1, 8, 133, 128]) - layer.4.k_cache: torch.Size([1, 8, 133, 128]) - layer.4.v_cache: torch.Size([1, 8, 133, 128]) - layer.4.output: torch.Size([1, 133, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.388s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 133, 128]) - layer.0.v_cache: torch.Size([1, 8, 133, 128]) - layer.1.k_cache: torch.Size([1, 8, 133, 128]) - layer.1.v_cache: torch.Size([1, 8, 133, 128]) - layer.2.k_cache: torch.Size([1, 8, 133, 128]) - layer.2.v_cache: torch.Size([1, 8, 133, 128]) - layer.3.k_cache: torch.Size([1, 8, 133, 128]) - layer.3.v_cache: torch.Size([1, 8, 133, 128]) - layer.4.k_cache: torch.Size([1, 8, 133, 128]) - layer.4.v_cache: torch.Size([1, 8, 133, 128]) - layer.4.output: torch.Size([1, 133, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02254625 22.98933307 - layer.0.v_cache 0.00000028 0.00031640 - layer.1.k_cache 0.00324298 1.50311256 - layer.1.v_cache 0.00000087 0.00107001 - layer.2.k_cache 0.00114197 0.71504309 - layer.2.v_cache 0.00000110 0.00162453 - layer.3.k_cache 0.00134409 0.81811013 - layer.3.v_cache 0.00000211 0.00261736 - layer.4.k_cache 0.00375619 1.59532005 - layer.4.v_cache 0.00000306 0.00434100 - layer.4.output 0.00014515 0.10563296 - ------------------------------------------------------------------------------------- - TOTAL 0.00232996 2.00381572 - (elements=1,906,688) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1906688 -Total Bytes 297384 -BPFP 1.2478 bits/point -EBPFP 2.4955 equivalent bits/point -MSE 2.003816 ----------------------- -------------------------------------------------------- -Time: 0.648s Load: 0.008s, Pack+Encode: 0.251s, Decode+Unpack: 0.388s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 133, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 133, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0038 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample52-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample52-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample53-layer4-item1.zst (70/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample53-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 125, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 125, 128) -Output shape: (1, 125, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.0.v_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.1.k_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.1.v_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.2.k_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.2.v_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.3.k_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.3.v_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.4.k_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.4.v_cache: torch.Size([1, 8, 125, 128]) -> torch.Size([1, 1, 125, 1024]) - layer.4.output: torch.Size([1, 125, 4096]) -> torch.Size([1, 1, 125, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,548B, BPFP=0.2218 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,924B, BPFP=1.3077 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,704B, BPFP=0.6690 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,020B, BPFP=1.4387 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,448B, BPFP=0.8405 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 23,164B, BPFP=1.4478 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,272B, BPFP=0.8920 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,016B, BPFP=1.3760 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,400B, BPFP=0.6500 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,468B, BPFP=1.4667 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 54,852B, BPFP=0.8571 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 125, 128]) - layer.0.v_cache: torch.Size([1, 8, 125, 128]) - layer.1.k_cache: torch.Size([1, 8, 125, 128]) - layer.1.v_cache: torch.Size([1, 8, 125, 128]) - layer.2.k_cache: torch.Size([1, 8, 125, 128]) - layer.2.v_cache: torch.Size([1, 8, 125, 128]) - layer.3.k_cache: torch.Size([1, 8, 125, 128]) - layer.3.v_cache: torch.Size([1, 8, 125, 128]) - layer.4.k_cache: torch.Size([1, 8, 125, 128]) - layer.4.v_cache: torch.Size([1, 8, 125, 128]) - layer.4.output: torch.Size([1, 125, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.291s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 125, 128]) - layer.0.v_cache: torch.Size([1, 8, 125, 128]) - layer.1.k_cache: torch.Size([1, 8, 125, 128]) - layer.1.v_cache: torch.Size([1, 8, 125, 128]) - layer.2.k_cache: torch.Size([1, 8, 125, 128]) - layer.2.v_cache: torch.Size([1, 8, 125, 128]) - layer.3.k_cache: torch.Size([1, 8, 125, 128]) - layer.3.v_cache: torch.Size([1, 8, 125, 128]) - layer.4.k_cache: torch.Size([1, 8, 125, 128]) - layer.4.v_cache: torch.Size([1, 8, 125, 128]) - layer.4.output: torch.Size([1, 125, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02456201 19.69818750 - layer.0.v_cache 0.00000027 0.00029687 - layer.1.k_cache 0.00318185 1.28760498 - layer.1.v_cache 0.00000072 0.00102448 - layer.2.k_cache 0.00113729 0.67525476 - layer.2.v_cache 0.00000107 0.00150017 - layer.3.k_cache 0.00136705 0.77711755 - layer.3.v_cache 0.00000203 0.00251005 - layer.4.k_cache 0.00364986 1.51643188 - layer.4.v_cache 0.00000297 0.00392962 - layer.4.output 0.00019454 0.10255595 - ------------------------------------------------------------------------------------- - TOTAL 0.00247738 1.74100584 - (elements=1,792,000) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1792000 -Total Bytes 219816 -BPFP 0.9813 bits/point -EBPFP 1.9626 equivalent bits/point -MSE 1.741006 ----------------------- -------------------------------------------------------- -Time: 0.504s Load: 0.007s, Pack+Encode: 0.206s, Decode+Unpack: 0.291s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 125, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 125, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7410 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample53-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample53-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample543-layer4-item1.zst (71/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample543-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 71, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 71, 128) -Output shape: (1, 71, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.0.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.1.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.1.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.2.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.2.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.3.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.3.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.4.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.4.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.4.output: torch.Size([1, 71, 4096]) -> torch.Size([1, 1, 71, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,420B, BPFP=0.2663 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,828B, BPFP=1.5216 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,676B, BPFP=0.8446 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,508B, BPFP=1.7064 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,268B, BPFP=1.1298 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,888B, BPFP=1.8583 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,308B, BPFP=1.1342 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,948B, BPFP=1.7548 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,896B, BPFP=0.7588 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 16,044B, BPFP=1.7654 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 36,776B, BPFP=1.0117 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 71, 128]) - layer.0.v_cache: torch.Size([1, 8, 71, 128]) - layer.1.k_cache: torch.Size([1, 8, 71, 128]) - layer.1.v_cache: torch.Size([1, 8, 71, 128]) - layer.2.k_cache: torch.Size([1, 8, 71, 128]) - layer.2.v_cache: torch.Size([1, 8, 71, 128]) - layer.3.k_cache: torch.Size([1, 8, 71, 128]) - layer.3.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.k_cache: torch.Size([1, 8, 71, 128]) - layer.4.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.output: torch.Size([1, 71, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.287s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 71, 128]) - layer.0.v_cache: torch.Size([1, 8, 71, 128]) - layer.1.k_cache: torch.Size([1, 8, 71, 128]) - layer.1.v_cache: torch.Size([1, 8, 71, 128]) - layer.2.k_cache: torch.Size([1, 8, 71, 128]) - layer.2.v_cache: torch.Size([1, 8, 71, 128]) - layer.3.k_cache: torch.Size([1, 8, 71, 128]) - layer.3.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.k_cache: torch.Size([1, 8, 71, 128]) - layer.4.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.output: torch.Size([1, 71, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02849298 31.65574796 - layer.0.v_cache 0.00000028 0.00031601 - layer.1.k_cache 0.00376352 1.82701218 - layer.1.v_cache 0.00000080 0.00118734 - layer.2.k_cache 0.00117123 0.78138335 - layer.2.v_cache 0.00000118 0.00164885 - layer.3.k_cache 0.00138233 0.90037343 - layer.3.v_cache 0.00000223 0.00276085 - layer.4.k_cache 0.00337686 1.73871365 - layer.4.v_cache 0.00000300 0.00448290 - layer.4.output 0.00019204 0.14598642 - ------------------------------------------------------------------------------------- - TOTAL 0.00278304 2.67839802 - (elements=1,017,856) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1017856 -Total Bytes 152560 -BPFP 1.1991 bits/point -EBPFP 2.3981 equivalent bits/point -MSE 2.678398 ----------------------- -------------------------------------------------------- -Time: 0.498s Load: 0.005s, Pack+Encode: 0.206s, Decode+Unpack: 0.287s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 71, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.6784 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample543-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample543-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample561-layer4-item1.zst (72/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample561-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 68, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 68, 128) -Output shape: (1, 68, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.0.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.1.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.1.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.2.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.2.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.3.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.3.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.4.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.4.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.4.output: torch.Size([1, 68, 4096]) -> torch.Size([1, 1, 68, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,512B, BPFP=0.2886 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,156B, BPFP=1.5115 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,488B, BPFP=0.8603 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 14,820B, BPFP=1.7027 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,296B, BPFP=1.1829 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 15,984B, BPFP=1.8364 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 9,940B, BPFP=1.1420 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,920B, BPFP=1.8290 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,784B, BPFP=0.7794 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 16,128B, BPFP=1.8529 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 33,936B, BPFP=0.9747 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 68, 128]) - layer.0.v_cache: torch.Size([1, 8, 68, 128]) - layer.1.k_cache: torch.Size([1, 8, 68, 128]) - layer.1.v_cache: torch.Size([1, 8, 68, 128]) - layer.2.k_cache: torch.Size([1, 8, 68, 128]) - layer.2.v_cache: torch.Size([1, 8, 68, 128]) - layer.3.k_cache: torch.Size([1, 8, 68, 128]) - layer.3.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.k_cache: torch.Size([1, 8, 68, 128]) - layer.4.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.output: torch.Size([1, 68, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 68, 128]) - layer.0.v_cache: torch.Size([1, 8, 68, 128]) - layer.1.k_cache: torch.Size([1, 8, 68, 128]) - layer.1.v_cache: torch.Size([1, 8, 68, 128]) - layer.2.k_cache: torch.Size([1, 8, 68, 128]) - layer.2.v_cache: torch.Size([1, 8, 68, 128]) - layer.3.k_cache: torch.Size([1, 8, 68, 128]) - layer.3.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.k_cache: torch.Size([1, 8, 68, 128]) - layer.4.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.output: torch.Size([1, 68, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02799314 29.11315200 - layer.0.v_cache 0.00000027 0.00032400 - layer.1.k_cache 0.00364027 1.90014267 - layer.1.v_cache 0.00000080 0.00127139 - layer.2.k_cache 0.00115812 0.80790699 - layer.2.v_cache 0.00000117 0.00178329 - layer.3.k_cache 0.00137295 0.91063286 - layer.3.v_cache 0.00000227 0.00305145 - layer.4.k_cache 0.00327663 1.85457274 - layer.4.v_cache 0.00000341 0.00510193 - layer.4.output 0.00018782 0.14440461 - ------------------------------------------------------------------------------------- - TOTAL 0.00272859 2.51253984 - (elements=974,848) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 974848 -Total Bytes 146964 -BPFP 1.2060 bits/point -EBPFP 2.4121 equivalent bits/point -MSE 2.512540 ----------------------- -------------------------------------------------------- -Time: 0.497s Load: 0.005s, Pack+Encode: 0.204s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 68, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5125 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample561-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample561-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample570-layer4-item1.zst (73/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample570-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 72, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 72, 128) -Output shape: (1, 72, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.0.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.1.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.1.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.2.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.2.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.3.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.3.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.4.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.4.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.4.output: torch.Size([1, 72, 4096]) -> torch.Size([1, 1, 72, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,436B, BPFP=0.2643 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,696B, BPFP=1.4861 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,040B, BPFP=0.8724 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,320B, BPFP=1.6623 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,084B, BPFP=1.2027 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,164B, BPFP=1.7539 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,744B, BPFP=1.1658 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,656B, BPFP=1.6988 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,180B, BPFP=0.7791 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 15,520B, BPFP=1.6840 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 31,828B, BPFP=0.8634 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 72, 128]) - layer.0.v_cache: torch.Size([1, 8, 72, 128]) - layer.1.k_cache: torch.Size([1, 8, 72, 128]) - layer.1.v_cache: torch.Size([1, 8, 72, 128]) - layer.2.k_cache: torch.Size([1, 8, 72, 128]) - layer.2.v_cache: torch.Size([1, 8, 72, 128]) - layer.3.k_cache: torch.Size([1, 8, 72, 128]) - layer.3.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.k_cache: torch.Size([1, 8, 72, 128]) - layer.4.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.output: torch.Size([1, 72, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.286s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 72, 128]) - layer.0.v_cache: torch.Size([1, 8, 72, 128]) - layer.1.k_cache: torch.Size([1, 8, 72, 128]) - layer.1.v_cache: torch.Size([1, 8, 72, 128]) - layer.2.k_cache: torch.Size([1, 8, 72, 128]) - layer.2.v_cache: torch.Size([1, 8, 72, 128]) - layer.3.k_cache: torch.Size([1, 8, 72, 128]) - layer.3.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.k_cache: torch.Size([1, 8, 72, 128]) - layer.4.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.output: torch.Size([1, 72, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02695644 30.13381619 - layer.0.v_cache 0.00000028 0.00032966 - layer.1.k_cache 0.00362044 1.83614180 - layer.1.v_cache 0.00000088 0.00118388 - layer.2.k_cache 0.00116887 0.79424148 - layer.2.v_cache 0.00000122 0.00173517 - layer.3.k_cache 0.00138034 0.92825635 - layer.3.v_cache 0.00000238 0.00285413 - layer.4.k_cache 0.00335768 1.77714740 - layer.4.v_cache 0.00000294 0.00441929 - layer.4.output 0.00021991 0.13757449 - ------------------------------------------------------------------------------------- - TOTAL 0.00266936 2.57360166 - (elements=1,032,192) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1032192 -Total Bytes 147668 -BPFP 1.1445 bits/point -EBPFP 2.2890 equivalent bits/point -MSE 2.573602 ----------------------- -------------------------------------------------------- -Time: 0.496s Load: 0.005s, Pack+Encode: 0.204s, Decode+Unpack: 0.286s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 72, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5736 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample570-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample570-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample581-layer4-item1.zst (74/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample581-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 73, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 73, 128) -Output shape: (1, 73, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.output: torch.Size([1, 73, 4096]) -> torch.Size([1, 1, 73, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,412B, BPFP=0.2581 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 14,112B, BPFP=1.5103 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,820B, BPFP=0.8369 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,172B, BPFP=1.6237 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,596B, BPFP=1.1340 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 15,756B, BPFP=1.6862 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,096B, BPFP=1.0805 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,820B, BPFP=1.6931 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,192B, BPFP=0.7697 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 15,932B, BPFP=1.7051 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 30,816B, BPFP=0.8245 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.output: torch.Size([1, 73, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.286s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.output: torch.Size([1, 73, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02610407 34.87606686 - layer.0.v_cache 0.00000028 0.00031853 - layer.1.k_cache 0.00357364 1.82362888 - layer.1.v_cache 0.00000075 0.00110866 - layer.2.k_cache 0.00114681 0.78460108 - layer.2.v_cache 0.00000108 0.00156639 - layer.3.k_cache 0.00136272 0.89380311 - layer.3.v_cache 0.00000212 0.00257523 - layer.4.k_cache 0.00341161 1.74794968 - layer.4.v_cache 0.00000293 0.00428179 - layer.4.output 0.00018354 0.13962334 - ------------------------------------------------------------------------------------- - TOTAL 0.00259573 2.90674240 - (elements=1,046,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1046528 -Total Bytes 145724 -BPFP 1.1140 bits/point -EBPFP 2.2279 equivalent bits/point -MSE 2.906742 ----------------------- -------------------------------------------------------- -Time: 0.496s Load: 0.006s, Pack+Encode: 0.204s, Decode+Unpack: 0.286s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 73, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.9067 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample581-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample581-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample584-layer4-item1.zst (75/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample584-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 72, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 72, 128) -Output shape: (1, 72, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.0.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.1.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.1.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.2.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.2.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.3.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.3.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.4.k_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.4.v_cache: torch.Size([1, 8, 72, 128]) -> torch.Size([1, 1, 72, 1024]) - layer.4.output: torch.Size([1, 72, 4096]) -> torch.Size([1, 1, 72, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,412B, BPFP=0.2617 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 14,124B, BPFP=1.5326 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,768B, BPFP=0.8429 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,676B, BPFP=1.7010 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,452B, BPFP=1.1341 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,972B, BPFP=1.8416 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,416B, BPFP=1.1302 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 16,748B, BPFP=1.8173 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,848B, BPFP=0.7431 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 16,984B, BPFP=1.8429 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 31,540B, BPFP=0.8556 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 72, 128]) - layer.0.v_cache: torch.Size([1, 8, 72, 128]) - layer.1.k_cache: torch.Size([1, 8, 72, 128]) - layer.1.v_cache: torch.Size([1, 8, 72, 128]) - layer.2.k_cache: torch.Size([1, 8, 72, 128]) - layer.2.v_cache: torch.Size([1, 8, 72, 128]) - layer.3.k_cache: torch.Size([1, 8, 72, 128]) - layer.3.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.k_cache: torch.Size([1, 8, 72, 128]) - layer.4.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.output: torch.Size([1, 72, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 72, 128]) - layer.0.v_cache: torch.Size([1, 8, 72, 128]) - layer.1.k_cache: torch.Size([1, 8, 72, 128]) - layer.1.v_cache: torch.Size([1, 8, 72, 128]) - layer.2.k_cache: torch.Size([1, 8, 72, 128]) - layer.2.v_cache: torch.Size([1, 8, 72, 128]) - layer.3.k_cache: torch.Size([1, 8, 72, 128]) - layer.3.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.k_cache: torch.Size([1, 8, 72, 128]) - layer.4.v_cache: torch.Size([1, 8, 72, 128]) - layer.4.output: torch.Size([1, 72, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02727311 30.10779148 - layer.0.v_cache 0.00000029 0.00031850 - layer.1.k_cache 0.00335621 1.79513020 - layer.1.v_cache 0.00000072 0.00102915 - layer.2.k_cache 0.00111533 0.75870159 - layer.2.v_cache 0.00000112 0.00152241 - layer.3.k_cache 0.00138591 0.86781120 - layer.3.v_cache 0.00000206 0.00246480 - layer.4.k_cache 0.00323834 1.81048033 - layer.4.v_cache 0.00000293 0.00412627 - layer.4.output 0.00022510 0.13510070 - ------------------------------------------------------------------------------------- - TOTAL 0.00266260 2.56355562 - (elements=1,032,192) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1032192 -Total Bytes 149940 -BPFP 1.1621 bits/point -EBPFP 2.3242 equivalent bits/point -MSE 2.563556 ----------------------- -------------------------------------------------------- -Time: 0.498s Load: 0.007s, Pack+Encode: 0.204s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 72, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 72, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5636 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample584-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample584-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample60-layer4-item1.zst (76/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample60-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 115, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 115, 128) -Output shape: (1, 115, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.0.v_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.1.k_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.1.v_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.2.k_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.2.v_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.3.k_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.3.v_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.4.k_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.4.v_cache: torch.Size([1, 8, 115, 128]) -> torch.Size([1, 1, 115, 1024]) - layer.4.output: torch.Size([1, 115, 4096]) -> torch.Size([1, 1, 115, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,396B, BPFP=0.2307 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,140B, BPFP=1.4361 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,160B, BPFP=0.6902 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,856B, BPFP=1.5527 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,752B, BPFP=0.8663 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,924B, BPFP=1.5573 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,536B, BPFP=0.9196 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,024B, BPFP=1.4962 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,564B, BPFP=0.7177 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,340B, BPFP=1.5856 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 66,384B, BPFP=1.1274 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 115, 128]) - layer.0.v_cache: torch.Size([1, 8, 115, 128]) - layer.1.k_cache: torch.Size([1, 8, 115, 128]) - layer.1.v_cache: torch.Size([1, 8, 115, 128]) - layer.2.k_cache: torch.Size([1, 8, 115, 128]) - layer.2.v_cache: torch.Size([1, 8, 115, 128]) - layer.3.k_cache: torch.Size([1, 8, 115, 128]) - layer.3.v_cache: torch.Size([1, 8, 115, 128]) - layer.4.k_cache: torch.Size([1, 8, 115, 128]) - layer.4.v_cache: torch.Size([1, 8, 115, 128]) - layer.4.output: torch.Size([1, 115, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.290s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 115, 128]) - layer.0.v_cache: torch.Size([1, 8, 115, 128]) - layer.1.k_cache: torch.Size([1, 8, 115, 128]) - layer.1.v_cache: torch.Size([1, 8, 115, 128]) - layer.2.k_cache: torch.Size([1, 8, 115, 128]) - layer.2.v_cache: torch.Size([1, 8, 115, 128]) - layer.3.k_cache: torch.Size([1, 8, 115, 128]) - layer.3.v_cache: torch.Size([1, 8, 115, 128]) - layer.4.k_cache: torch.Size([1, 8, 115, 128]) - layer.4.v_cache: torch.Size([1, 8, 115, 128]) - layer.4.output: torch.Size([1, 115, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02486876 21.03937882 - layer.0.v_cache 0.00000027 0.00031935 - layer.1.k_cache 0.00345256 1.50612448 - layer.1.v_cache 0.00000081 0.00116631 - layer.2.k_cache 0.00114556 0.74852335 - layer.2.v_cache 0.00000111 0.00167164 - layer.3.k_cache 0.00139466 0.83308013 - layer.3.v_cache 0.00000215 0.00278080 - layer.4.k_cache 0.00339940 1.63085884 - layer.4.v_cache 0.00000318 0.00475511 - layer.4.output 0.00018696 0.12124232 - ------------------------------------------------------------------------------------- - TOTAL 0.00250116 1.87525915 - (elements=1,648,640) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1648640 -Total Bytes 229076 -BPFP 1.1116 bits/point -EBPFP 2.2232 equivalent bits/point -MSE 1.875259 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.008s, Pack+Encode: 0.204s, Decode+Unpack: 0.290s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 115, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 115, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8753 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample60-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample60-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample600-layer4-item1.zst (77/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample600-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 71, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 71, 128) -Output shape: (1, 71, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.0.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.1.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.1.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.2.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.2.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.3.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.3.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.4.k_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.4.v_cache: torch.Size([1, 8, 71, 128]) -> torch.Size([1, 1, 71, 1024]) - layer.4.output: torch.Size([1, 71, 4096]) -> torch.Size([1, 1, 71, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,432B, BPFP=0.2676 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,528B, BPFP=1.4886 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,684B, BPFP=0.8455 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,432B, BPFP=1.6981 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,208B, BPFP=1.1232 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,376B, BPFP=1.8019 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,236B, BPFP=1.1263 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,788B, BPFP=1.7372 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,800B, BPFP=0.7482 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 15,672B, BPFP=1.7245 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 32,212B, BPFP=0.8861 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 71, 128]) - layer.0.v_cache: torch.Size([1, 8, 71, 128]) - layer.1.k_cache: torch.Size([1, 8, 71, 128]) - layer.1.v_cache: torch.Size([1, 8, 71, 128]) - layer.2.k_cache: torch.Size([1, 8, 71, 128]) - layer.2.v_cache: torch.Size([1, 8, 71, 128]) - layer.3.k_cache: torch.Size([1, 8, 71, 128]) - layer.3.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.k_cache: torch.Size([1, 8, 71, 128]) - layer.4.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.output: torch.Size([1, 71, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.292s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 71, 128]) - layer.0.v_cache: torch.Size([1, 8, 71, 128]) - layer.1.k_cache: torch.Size([1, 8, 71, 128]) - layer.1.v_cache: torch.Size([1, 8, 71, 128]) - layer.2.k_cache: torch.Size([1, 8, 71, 128]) - layer.2.v_cache: torch.Size([1, 8, 71, 128]) - layer.3.k_cache: torch.Size([1, 8, 71, 128]) - layer.3.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.k_cache: torch.Size([1, 8, 71, 128]) - layer.4.v_cache: torch.Size([1, 8, 71, 128]) - layer.4.output: torch.Size([1, 71, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02663550 32.13707293 - layer.0.v_cache 0.00000027 0.00031943 - layer.1.k_cache 0.00355294 1.98578182 - layer.1.v_cache 0.00000089 0.00120002 - layer.2.k_cache 0.00117088 0.81865187 - layer.2.v_cache 0.00000116 0.00168926 - layer.3.k_cache 0.00139091 0.92279536 - layer.3.v_cache 0.00000218 0.00274574 - layer.4.k_cache 0.00325425 1.89272147 - layer.4.v_cache 0.00000306 0.00440514 - layer.4.output 0.00022008 0.15061976 - ------------------------------------------------------------------------------------- - TOTAL 0.00263517 2.74070443 - (elements=1,017,856) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1017856 -Total Bytes 146368 -BPFP 1.1504 bits/point -EBPFP 2.3008 equivalent bits/point -MSE 2.740704 ----------------------- -------------------------------------------------------- -Time: 0.502s Load: 0.005s, Pack+Encode: 0.206s, Decode+Unpack: 0.292s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 71, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 71, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.7407 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample600-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample600-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample622-layer4-item1.zst (78/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample622-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 73, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 73, 128) -Output shape: (1, 73, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.output: torch.Size([1, 73, 4096]) -> torch.Size([1, 1, 73, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,428B, BPFP=0.2598 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 14,664B, BPFP=1.5693 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,200B, BPFP=0.8776 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,776B, BPFP=1.6884 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,612B, BPFP=1.1357 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,208B, BPFP=1.7346 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,228B, BPFP=1.0946 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 16,280B, BPFP=1.7423 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,808B, BPFP=0.7286 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 16,300B, BPFP=1.7444 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 33,448B, BPFP=0.8949 -⌛️ [2/4] FRONTEND: Frontend time: 0.212s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.output: torch.Size([1, 73, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.285s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.output: torch.Size([1, 73, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02450750 31.69682082 - layer.0.v_cache 0.00000028 0.00033462 - layer.1.k_cache 0.00347943 1.89593255 - layer.1.v_cache 0.00000079 0.00120025 - layer.2.k_cache 0.00114165 0.80074373 - layer.2.v_cache 0.00000118 0.00172880 - layer.3.k_cache 0.00136377 0.92796409 - layer.3.v_cache 0.00000212 0.00277265 - layer.4.k_cache 0.00321585 1.81678856 - layer.4.v_cache 0.00000299 0.00450741 - layer.4.output 0.00023923 0.15772991 - ------------------------------------------------------------------------------------- - TOTAL 0.00247661 2.69855094 - (elements=1,046,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1046528 -Total Bytes 150952 -BPFP 1.1539 bits/point -EBPFP 2.3079 equivalent bits/point -MSE 2.698551 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.005s, Pack+Encode: 0.212s, Decode+Unpack: 0.285s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 73, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.6986 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample622-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample622-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample65-layer4-item1.zst (79/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample65-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 112, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 112, 128) -Output shape: (1, 112, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.0.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.1.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.1.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.2.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.2.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.3.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.3.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.4.k_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.4.v_cache: torch.Size([1, 8, 112, 128]) -> torch.Size([1, 1, 112, 1024]) - layer.4.output: torch.Size([1, 112, 4096]) -> torch.Size([1, 1, 112, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,428B, BPFP=0.2391 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,760B, BPFP=1.5179 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,532B, BPFP=0.7347 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,012B, BPFP=1.6052 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,132B, BPFP=0.9160 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,888B, BPFP=1.5965 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,712B, BPFP=0.9565 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,844B, BPFP=1.5237 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,280B, BPFP=0.7868 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,012B, BPFP=1.6052 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 66,912B, BPFP=1.1669 -⌛️ [2/4] FRONTEND: Frontend time: 0.201s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 112, 128]) - layer.0.v_cache: torch.Size([1, 8, 112, 128]) - layer.1.k_cache: torch.Size([1, 8, 112, 128]) - layer.1.v_cache: torch.Size([1, 8, 112, 128]) - layer.2.k_cache: torch.Size([1, 8, 112, 128]) - layer.2.v_cache: torch.Size([1, 8, 112, 128]) - layer.3.k_cache: torch.Size([1, 8, 112, 128]) - layer.3.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.k_cache: torch.Size([1, 8, 112, 128]) - layer.4.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.output: torch.Size([1, 112, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.278s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 112, 128]) - layer.0.v_cache: torch.Size([1, 8, 112, 128]) - layer.1.k_cache: torch.Size([1, 8, 112, 128]) - layer.1.v_cache: torch.Size([1, 8, 112, 128]) - layer.2.k_cache: torch.Size([1, 8, 112, 128]) - layer.2.v_cache: torch.Size([1, 8, 112, 128]) - layer.3.k_cache: torch.Size([1, 8, 112, 128]) - layer.3.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.k_cache: torch.Size([1, 8, 112, 128]) - layer.4.v_cache: torch.Size([1, 8, 112, 128]) - layer.4.output: torch.Size([1, 112, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02850304 22.86221967 - layer.0.v_cache 0.00000027 0.00033428 - layer.1.k_cache 0.00329348 1.52740329 - layer.1.v_cache 0.00000088 0.00130574 - layer.2.k_cache 0.00114370 0.73012652 - layer.2.v_cache 0.00000135 0.00174696 - layer.3.k_cache 0.00133332 0.83113439 - layer.3.v_cache 0.00000257 0.00307199 - layer.4.k_cache 0.00339413 1.56445530 - layer.4.v_cache 0.00000321 0.00469212 - layer.4.output 0.00019581 0.11375890 - ------------------------------------------------------------------------------------- - TOTAL 0.00274709 1.99868042 - (elements=1,605,632) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1605632 -Total Bytes 231512 -BPFP 1.1535 bits/point -EBPFP 2.3070 equivalent bits/point -MSE 1.998680 ----------------------- -------------------------------------------------------- -Time: 0.487s Load: 0.007s, Pack+Encode: 0.201s, Decode+Unpack: 0.278s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 112, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 112, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9987 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample65-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample65-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample656-layer4-item1.zst (80/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample656-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 68, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 68, 128) -Output shape: (1, 68, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.0.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.1.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.1.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.2.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.2.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.3.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.3.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.4.k_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.4.v_cache: torch.Size([1, 8, 68, 128]) -> torch.Size([1, 1, 68, 1024]) - layer.4.output: torch.Size([1, 68, 4096]) -> torch.Size([1, 1, 68, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,488B, BPFP=0.2858 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,136B, BPFP=1.5092 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,224B, BPFP=0.8300 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 14,836B, BPFP=1.7045 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,184B, BPFP=1.1700 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 15,720B, BPFP=1.8061 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 9,868B, BPFP=1.1337 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,396B, BPFP=1.7688 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,464B, BPFP=0.7426 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 15,140B, BPFP=1.7394 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 32,032B, BPFP=0.9200 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 68, 128]) - layer.0.v_cache: torch.Size([1, 8, 68, 128]) - layer.1.k_cache: torch.Size([1, 8, 68, 128]) - layer.1.v_cache: torch.Size([1, 8, 68, 128]) - layer.2.k_cache: torch.Size([1, 8, 68, 128]) - layer.2.v_cache: torch.Size([1, 8, 68, 128]) - layer.3.k_cache: torch.Size([1, 8, 68, 128]) - layer.3.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.k_cache: torch.Size([1, 8, 68, 128]) - layer.4.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.output: torch.Size([1, 68, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 68, 128]) - layer.0.v_cache: torch.Size([1, 8, 68, 128]) - layer.1.k_cache: torch.Size([1, 8, 68, 128]) - layer.1.v_cache: torch.Size([1, 8, 68, 128]) - layer.2.k_cache: torch.Size([1, 8, 68, 128]) - layer.2.v_cache: torch.Size([1, 8, 68, 128]) - layer.3.k_cache: torch.Size([1, 8, 68, 128]) - layer.3.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.k_cache: torch.Size([1, 8, 68, 128]) - layer.4.v_cache: torch.Size([1, 8, 68, 128]) - layer.4.output: torch.Size([1, 68, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02721363 29.31250718 - layer.0.v_cache 0.00000029 0.00031982 - layer.1.k_cache 0.00369932 1.81951590 - layer.1.v_cache 0.00000081 0.00113740 - layer.2.k_cache 0.00117793 0.78523799 - layer.2.v_cache 0.00000112 0.00165766 - layer.3.k_cache 0.00140028 0.91972699 - layer.3.v_cache 0.00000224 0.00283919 - layer.4.k_cache 0.00315349 1.79281482 - layer.4.v_cache 0.00000301 0.00442506 - layer.4.output 0.00021496 0.13864892 - ------------------------------------------------------------------------------------- - TOTAL 0.00267943 2.51391269 - (elements=974,848) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 974848 -Total Bytes 142488 -BPFP 1.1693 bits/point -EBPFP 2.3386 equivalent bits/point -MSE 2.513913 ----------------------- -------------------------------------------------------- -Time: 0.481s Load: 0.005s, Pack+Encode: 0.200s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 68, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 68, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5139 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample656-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample656-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample66-layer4-item1.zst (81/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample66-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 136, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 136, 128) -Output shape: (1, 136, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.0.v_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.1.k_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.1.v_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.2.k_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.2.v_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.3.k_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.3.v_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.4.k_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.4.v_cache: torch.Size([1, 8, 136, 128]) -> torch.Size([1, 1, 136, 1024]) - layer.4.output: torch.Size([1, 136, 4096]) -> torch.Size([1, 1, 136, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,896B, BPFP=0.2812 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 27,752B, BPFP=1.5942 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 15,812B, BPFP=0.9083 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 29,580B, BPFP=1.6992 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,440B, BPFP=1.0593 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 29,732B, BPFP=1.7080 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,920B, BPFP=1.1443 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 29,812B, BPFP=1.7125 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 16,636B, BPFP=0.9557 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 30,412B, BPFP=1.7470 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 97,984B, BPFP=1.4072 -⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 136, 128]) - layer.0.v_cache: torch.Size([1, 8, 136, 128]) - layer.1.k_cache: torch.Size([1, 8, 136, 128]) - layer.1.v_cache: torch.Size([1, 8, 136, 128]) - layer.2.k_cache: torch.Size([1, 8, 136, 128]) - layer.2.v_cache: torch.Size([1, 8, 136, 128]) - layer.3.k_cache: torch.Size([1, 8, 136, 128]) - layer.3.v_cache: torch.Size([1, 8, 136, 128]) - layer.4.k_cache: torch.Size([1, 8, 136, 128]) - layer.4.v_cache: torch.Size([1, 8, 136, 128]) - layer.4.output: torch.Size([1, 136, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.375s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 136, 128]) - layer.0.v_cache: torch.Size([1, 8, 136, 128]) - layer.1.k_cache: torch.Size([1, 8, 136, 128]) - layer.1.v_cache: torch.Size([1, 8, 136, 128]) - layer.2.k_cache: torch.Size([1, 8, 136, 128]) - layer.2.v_cache: torch.Size([1, 8, 136, 128]) - layer.3.k_cache: torch.Size([1, 8, 136, 128]) - layer.3.v_cache: torch.Size([1, 8, 136, 128]) - layer.4.k_cache: torch.Size([1, 8, 136, 128]) - layer.4.v_cache: torch.Size([1, 8, 136, 128]) - layer.4.output: torch.Size([1, 136, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02725438 24.13402602 - layer.0.v_cache 0.00000027 0.00031222 - layer.1.k_cache 0.00321745 1.54895536 - layer.1.v_cache 0.00000084 0.00106397 - layer.2.k_cache 0.00114365 0.72213106 - layer.2.v_cache 0.00000128 0.00153924 - layer.3.k_cache 0.00137918 0.80816168 - layer.3.v_cache 0.00000219 0.00256459 - layer.4.k_cache 0.00340721 1.64662170 - layer.4.v_cache 0.00000338 0.00430364 - layer.4.output 0.00019585 0.12502809 - ------------------------------------------------------------------------------------- - TOTAL 0.00265666 2.09784227 - (elements=1,949,696) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1949696 -Total Bytes 320976 -BPFP 1.3170 bits/point -EBPFP 2.6341 equivalent bits/point -MSE 2.097842 ----------------------- -------------------------------------------------------- -Time: 0.632s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.375s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 136, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 136, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0978 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample66-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample66-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample663-layer4-item1.zst (82/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample663-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 67, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.004s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 67, 128) -Output shape: (1, 67, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.0.v_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.1.k_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.1.v_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.2.k_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.2.v_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.3.k_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.3.v_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.4.k_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.4.v_cache: torch.Size([1, 8, 67, 128]) -> torch.Size([1, 1, 67, 1024]) - layer.4.output: torch.Size([1, 67, 4096]) -> torch.Size([1, 1, 67, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,456B, BPFP=0.2864 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,036B, BPFP=1.5201 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,056B, BPFP=0.8228 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 14,972B, BPFP=1.7458 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 9,416B, BPFP=1.0979 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,328B, BPFP=1.9039 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 9,436B, BPFP=1.1003 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 16,048B, BPFP=1.8713 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,060B, BPFP=0.7066 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 15,840B, BPFP=1.8470 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 32,808B, BPFP=0.9564 -⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 67, 128]) - layer.0.v_cache: torch.Size([1, 8, 67, 128]) - layer.1.k_cache: torch.Size([1, 8, 67, 128]) - layer.1.v_cache: torch.Size([1, 8, 67, 128]) - layer.2.k_cache: torch.Size([1, 8, 67, 128]) - layer.2.v_cache: torch.Size([1, 8, 67, 128]) - layer.3.k_cache: torch.Size([1, 8, 67, 128]) - layer.3.v_cache: torch.Size([1, 8, 67, 128]) - layer.4.k_cache: torch.Size([1, 8, 67, 128]) - layer.4.v_cache: torch.Size([1, 8, 67, 128]) - layer.4.output: torch.Size([1, 67, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 67, 128]) - layer.0.v_cache: torch.Size([1, 8, 67, 128]) - layer.1.k_cache: torch.Size([1, 8, 67, 128]) - layer.1.v_cache: torch.Size([1, 8, 67, 128]) - layer.2.k_cache: torch.Size([1, 8, 67, 128]) - layer.2.v_cache: torch.Size([1, 8, 67, 128]) - layer.3.k_cache: torch.Size([1, 8, 67, 128]) - layer.3.v_cache: torch.Size([1, 8, 67, 128]) - layer.4.k_cache: torch.Size([1, 8, 67, 128]) - layer.4.v_cache: torch.Size([1, 8, 67, 128]) - layer.4.output: torch.Size([1, 67, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02769006 29.24429731 - layer.0.v_cache 0.00000028 0.00033600 - layer.1.k_cache 0.00368886 1.76873244 - layer.1.v_cache 0.00000112 0.00125036 - layer.2.k_cache 0.00116499 0.81093256 - layer.2.v_cache 0.00000120 0.00182768 - layer.3.k_cache 0.00139456 0.92811220 - layer.3.v_cache 0.00000232 0.00305506 - layer.4.k_cache 0.00321227 1.85703471 - layer.4.v_cache 0.00000307 0.00466780 - layer.4.output 0.00021890 0.14411801 - ------------------------------------------------------------------------------------- - TOTAL 0.00271674 2.51405130 - (elements=960,512) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 960512 -Total Bytes 143456 -BPFP 1.1948 bits/point -EBPFP 2.3897 equivalent bits/point -MSE 2.514051 ----------------------- -------------------------------------------------------- -Time: 0.481s Load: 0.004s, Pack+Encode: 0.200s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 67, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 67, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5141 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample663-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample663-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample68-layer4-item1.zst (83/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample68-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 123, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 123, 128) -Output shape: (1, 123, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.0.v_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.1.k_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.1.v_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.2.k_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.2.v_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.3.k_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.3.v_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.4.k_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.4.v_cache: torch.Size([1, 8, 123, 128]) -> torch.Size([1, 1, 123, 1024]) - layer.4.output: torch.Size([1, 123, 4096]) -> torch.Size([1, 1, 123, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,356B, BPFP=0.2132 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,752B, BPFP=1.3816 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,408B, BPFP=0.6611 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,148B, BPFP=1.4703 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,872B, BPFP=0.8176 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 23,656B, BPFP=1.5025 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,116B, BPFP=0.8966 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,856B, BPFP=1.4517 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,420B, BPFP=0.6618 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 24,304B, BPFP=1.5437 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 60,984B, BPFP=0.9684 -⌛️ [2/4] FRONTEND: Frontend time: 0.198s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 123, 128]) - layer.0.v_cache: torch.Size([1, 8, 123, 128]) - layer.1.k_cache: torch.Size([1, 8, 123, 128]) - layer.1.v_cache: torch.Size([1, 8, 123, 128]) - layer.2.k_cache: torch.Size([1, 8, 123, 128]) - layer.2.v_cache: torch.Size([1, 8, 123, 128]) - layer.3.k_cache: torch.Size([1, 8, 123, 128]) - layer.3.v_cache: torch.Size([1, 8, 123, 128]) - layer.4.k_cache: torch.Size([1, 8, 123, 128]) - layer.4.v_cache: torch.Size([1, 8, 123, 128]) - layer.4.output: torch.Size([1, 123, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.280s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 123, 128]) - layer.0.v_cache: torch.Size([1, 8, 123, 128]) - layer.1.k_cache: torch.Size([1, 8, 123, 128]) - layer.1.v_cache: torch.Size([1, 8, 123, 128]) - layer.2.k_cache: torch.Size([1, 8, 123, 128]) - layer.2.v_cache: torch.Size([1, 8, 123, 128]) - layer.3.k_cache: torch.Size([1, 8, 123, 128]) - layer.3.v_cache: torch.Size([1, 8, 123, 128]) - layer.4.k_cache: torch.Size([1, 8, 123, 128]) - layer.4.v_cache: torch.Size([1, 8, 123, 128]) - layer.4.output: torch.Size([1, 123, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02602563 21.78852261 - layer.0.v_cache 0.00000027 0.00029320 - layer.1.k_cache 0.00336899 1.48416857 - layer.1.v_cache 0.00000079 0.00102121 - layer.2.k_cache 0.00113470 0.70552454 - layer.2.v_cache 0.00000113 0.00149773 - layer.3.k_cache 0.00144113 0.80815739 - layer.3.v_cache 0.00000264 0.00262014 - layer.4.k_cache 0.00356668 1.63400417 - layer.4.v_cache 0.00000295 0.00404206 - layer.4.output 0.00019508 0.12323019 - ------------------------------------------------------------------------------------- - TOTAL 0.00259466 1.92305517 - (elements=1,763,328) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1763328 -Total Bytes 227872 -BPFP 1.0338 bits/point -EBPFP 2.0677 equivalent bits/point -MSE 1.923055 ----------------------- -------------------------------------------------------- -Time: 0.486s Load: 0.007s, Pack+Encode: 0.198s, Decode+Unpack: 0.280s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 123, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 123, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9231 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample68-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample68-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample7-layer4-item1.zst (84/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample7-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 157, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 157, 128) -Output shape: (1, 157, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.output: torch.Size([1, 157, 4096]) -> torch.Size([1, 1, 157, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,996B, BPFP=0.2486 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 29,592B, BPFP=1.4725 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,940B, BPFP=0.7434 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 31,172B, BPFP=1.5512 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,140B, BPFP=0.9027 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 31,604B, BPFP=1.5727 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,180B, BPFP=0.9544 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 30,844B, BPFP=1.5348 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 15,176B, BPFP=0.7552 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 32,404B, BPFP=1.6125 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 90,712B, BPFP=1.1285 -⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.output: torch.Size([1, 157, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.376s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.output: torch.Size([1, 157, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02423319 20.33432959 - layer.0.v_cache 0.00000027 0.00030264 - layer.1.k_cache 0.00309445 1.49565659 - layer.1.v_cache 0.00000088 0.00107313 - layer.2.k_cache 0.00117908 0.70979275 - layer.2.v_cache 0.00000110 0.00153361 - layer.3.k_cache 0.00138553 0.79595194 - layer.3.v_cache 0.00000215 0.00266934 - layer.4.k_cache 0.00349636 1.57666764 - layer.4.v_cache 0.00000310 0.00417660 - layer.4.output 0.00020206 0.09964322 - ------------------------------------------------------------------------------------- - TOTAL 0.00244317 1.80862333 - (elements=2,250,752) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2250752 -Total Bytes 318760 -BPFP 1.1330 bits/point -EBPFP 2.2660 equivalent bits/point -MSE 1.808623 ----------------------- -------------------------------------------------------- -Time: 0.632s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.376s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 157, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8086 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample7-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample7-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample736-layer4-item1.zst (85/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample736-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 75, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 75, 128) -Output shape: (1, 75, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.0.v_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.1.k_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.1.v_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.2.k_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.2.v_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.3.k_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.3.v_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.4.k_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.4.v_cache: torch.Size([1, 8, 75, 128]) -> torch.Size([1, 1, 75, 1024]) - layer.4.output: torch.Size([1, 75, 4096]) -> torch.Size([1, 1, 75, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,496B, BPFP=0.2600 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 14,788B, BPFP=1.5404 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 8,256B, BPFP=0.8600 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 16,132B, BPFP=1.6804 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,948B, BPFP=1.1404 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,192B, BPFP=1.6867 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,492B, BPFP=1.0929 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 16,684B, BPFP=1.7379 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,256B, BPFP=0.7558 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 16,680B, BPFP=1.7375 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,716B, BPFP=1.0343 -⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 75, 128]) - layer.0.v_cache: torch.Size([1, 8, 75, 128]) - layer.1.k_cache: torch.Size([1, 8, 75, 128]) - layer.1.v_cache: torch.Size([1, 8, 75, 128]) - layer.2.k_cache: torch.Size([1, 8, 75, 128]) - layer.2.v_cache: torch.Size([1, 8, 75, 128]) - layer.3.k_cache: torch.Size([1, 8, 75, 128]) - layer.3.v_cache: torch.Size([1, 8, 75, 128]) - layer.4.k_cache: torch.Size([1, 8, 75, 128]) - layer.4.v_cache: torch.Size([1, 8, 75, 128]) - layer.4.output: torch.Size([1, 75, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.276s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 75, 128]) - layer.0.v_cache: torch.Size([1, 8, 75, 128]) - layer.1.k_cache: torch.Size([1, 8, 75, 128]) - layer.1.v_cache: torch.Size([1, 8, 75, 128]) - layer.2.k_cache: torch.Size([1, 8, 75, 128]) - layer.2.v_cache: torch.Size([1, 8, 75, 128]) - layer.3.k_cache: torch.Size([1, 8, 75, 128]) - layer.3.v_cache: torch.Size([1, 8, 75, 128]) - layer.4.k_cache: torch.Size([1, 8, 75, 128]) - layer.4.v_cache: torch.Size([1, 8, 75, 128]) - layer.4.output: torch.Size([1, 75, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02489581 30.69842448 - layer.0.v_cache 0.00000028 0.00033833 - layer.1.k_cache 0.00341811 1.83421163 - layer.1.v_cache 0.00000080 0.00114462 - layer.2.k_cache 0.00110890 0.77263062 - layer.2.v_cache 0.00000112 0.00161080 - layer.3.k_cache 0.00134674 0.89568481 - layer.3.v_cache 0.00000209 0.00268931 - layer.4.k_cache 0.00330252 1.74018941 - layer.4.v_cache 0.00000296 0.00443397 - layer.4.output 0.00020017 0.14408955 - ------------------------------------------------------------------------------------- - TOTAL 0.00249143 2.60912258 - (elements=1,075,200) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1075200 -Total Bytes 159640 -BPFP 1.1878 bits/point -EBPFP 2.3756 equivalent bits/point -MSE 2.609123 ----------------------- -------------------------------------------------------- -Time: 0.480s Load: 0.005s, Pack+Encode: 0.199s, Decode+Unpack: 0.276s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 75, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 75, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.6091 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample736-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample736-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample74-layer4-item1.zst (86/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample74-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 122, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 122, 128) -Output shape: (1, 122, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.0.v_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.1.k_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.1.v_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.2.k_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.2.v_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.3.k_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.3.v_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.4.k_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.4.v_cache: torch.Size([1, 8, 122, 128]) -> torch.Size([1, 1, 122, 1024]) - layer.4.output: torch.Size([1, 122, 4096]) -> torch.Size([1, 1, 122, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,304B, BPFP=0.2116 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,876B, BPFP=1.4009 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,588B, BPFP=0.6780 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,776B, BPFP=1.5225 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,268B, BPFP=0.8496 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 23,552B, BPFP=1.5082 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,224B, BPFP=0.9109 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,668B, BPFP=1.4516 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,348B, BPFP=0.6627 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,796B, BPFP=1.5238 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 56,304B, BPFP=0.9014 -⌛️ [2/4] FRONTEND: Frontend time: 0.209s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 122, 128]) - layer.0.v_cache: torch.Size([1, 8, 122, 128]) - layer.1.k_cache: torch.Size([1, 8, 122, 128]) - layer.1.v_cache: torch.Size([1, 8, 122, 128]) - layer.2.k_cache: torch.Size([1, 8, 122, 128]) - layer.2.v_cache: torch.Size([1, 8, 122, 128]) - layer.3.k_cache: torch.Size([1, 8, 122, 128]) - layer.3.v_cache: torch.Size([1, 8, 122, 128]) - layer.4.k_cache: torch.Size([1, 8, 122, 128]) - layer.4.v_cache: torch.Size([1, 8, 122, 128]) - layer.4.output: torch.Size([1, 122, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.293s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 122, 128]) - layer.0.v_cache: torch.Size([1, 8, 122, 128]) - layer.1.k_cache: torch.Size([1, 8, 122, 128]) - layer.1.v_cache: torch.Size([1, 8, 122, 128]) - layer.2.k_cache: torch.Size([1, 8, 122, 128]) - layer.2.v_cache: torch.Size([1, 8, 122, 128]) - layer.3.k_cache: torch.Size([1, 8, 122, 128]) - layer.3.v_cache: torch.Size([1, 8, 122, 128]) - layer.4.k_cache: torch.Size([1, 8, 122, 128]) - layer.4.v_cache: torch.Size([1, 8, 122, 128]) - layer.4.output: torch.Size([1, 122, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02525679 21.07366643 - layer.0.v_cache 0.00000028 0.00031546 - layer.1.k_cache 0.00316206 1.42776001 - layer.1.v_cache 0.00000097 0.00112225 - layer.2.k_cache 0.00113900 0.68271837 - layer.2.v_cache 0.00000117 0.00161244 - layer.3.k_cache 0.00136642 0.78437449 - layer.3.v_cache 0.00000221 0.00258200 - layer.4.k_cache 0.00339647 1.46843982 - layer.4.v_cache 0.00000316 0.00420270 - layer.4.output 0.00022994 0.11387927 - ------------------------------------------------------------------------------------- - TOTAL 0.00251773 1.85016508 - (elements=1,748,992) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1748992 -Total Bytes 223704 -BPFP 1.0232 bits/point -EBPFP 2.0465 equivalent bits/point -MSE 1.850165 ----------------------- -------------------------------------------------------- -Time: 0.511s Load: 0.008s, Pack+Encode: 0.209s, Decode+Unpack: 0.293s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 122, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 122, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8502 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample74-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample74-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample75-layer4-item1.zst (87/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample75-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 142, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 142, 128) -Output shape: (1, 142, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.0.v_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.1.k_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.1.v_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.2.k_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.2.v_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.3.k_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.3.v_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.4.k_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.4.v_cache: torch.Size([1, 8, 142, 128]) -> torch.Size([1, 1, 142, 1024]) - layer.4.output: torch.Size([1, 142, 4096]) -> torch.Size([1, 1, 142, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 5,048B, BPFP=0.2777 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 28,032B, BPFP=1.5423 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 15,828B, BPFP=0.8708 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 30,632B, BPFP=1.6853 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 18,380B, BPFP=1.0112 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 30,912B, BPFP=1.7007 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,640B, BPFP=1.0805 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 30,652B, BPFP=1.6864 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 16,656B, BPFP=0.9164 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 31,372B, BPFP=1.7260 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 101,896B, BPFP=1.4015 -⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 142, 128]) - layer.0.v_cache: torch.Size([1, 8, 142, 128]) - layer.1.k_cache: torch.Size([1, 8, 142, 128]) - layer.1.v_cache: torch.Size([1, 8, 142, 128]) - layer.2.k_cache: torch.Size([1, 8, 142, 128]) - layer.2.v_cache: torch.Size([1, 8, 142, 128]) - layer.3.k_cache: torch.Size([1, 8, 142, 128]) - layer.3.v_cache: torch.Size([1, 8, 142, 128]) - layer.4.k_cache: torch.Size([1, 8, 142, 128]) - layer.4.v_cache: torch.Size([1, 8, 142, 128]) - layer.4.output: torch.Size([1, 142, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.392s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 142, 128]) - layer.0.v_cache: torch.Size([1, 8, 142, 128]) - layer.1.k_cache: torch.Size([1, 8, 142, 128]) - layer.1.v_cache: torch.Size([1, 8, 142, 128]) - layer.2.k_cache: torch.Size([1, 8, 142, 128]) - layer.2.v_cache: torch.Size([1, 8, 142, 128]) - layer.3.k_cache: torch.Size([1, 8, 142, 128]) - layer.3.v_cache: torch.Size([1, 8, 142, 128]) - layer.4.k_cache: torch.Size([1, 8, 142, 128]) - layer.4.v_cache: torch.Size([1, 8, 142, 128]) - layer.4.output: torch.Size([1, 142, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02301751 23.40720937 - layer.0.v_cache 0.00000029 0.00030967 - layer.1.k_cache 0.00315333 1.50548930 - layer.1.v_cache 0.00000086 0.00107331 - layer.2.k_cache 0.00112591 0.69633269 - layer.2.v_cache 0.00000130 0.00153525 - layer.3.k_cache 0.00130425 0.79467687 - layer.3.v_cache 0.00000237 0.00276580 - layer.4.k_cache 0.00348660 1.61720093 - layer.4.v_cache 0.00000314 0.00416674 - layer.4.output 0.00018111 0.12054395 - ------------------------------------------------------------------------------------- - TOTAL 0.00234429 2.03663827 - (elements=2,035,712) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2035712 -Total Bytes 329048 -BPFP 1.2931 bits/point -EBPFP 2.5862 equivalent bits/point -MSE 2.036638 ----------------------- -------------------------------------------------------- -Time: 0.657s Load: 0.008s, Pack+Encode: 0.258s, Decode+Unpack: 0.392s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 142, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 142, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.0366 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample75-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample75-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample777-layer4-item1.zst (88/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample777-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 74, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 74, 128) -Output shape: (1, 74, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.0.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.1.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.1.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.2.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.2.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.3.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.3.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.4.k_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.4.v_cache: torch.Size([1, 8, 74, 128]) -> torch.Size([1, 1, 74, 1024]) - layer.4.output: torch.Size([1, 74, 4096]) -> torch.Size([1, 1, 74, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,416B, BPFP=0.2551 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 15,240B, BPFP=1.6090 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,904B, BPFP=0.8345 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 17,040B, BPFP=1.7990 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,092B, BPFP=1.1710 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 17,400B, BPFP=1.8370 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,480B, BPFP=1.1064 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 17,048B, BPFP=1.7998 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 7,188B, BPFP=0.7589 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 17,820B, BPFP=1.8813 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 30,988B, BPFP=0.8179 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 74, 128]) - layer.0.v_cache: torch.Size([1, 8, 74, 128]) - layer.1.k_cache: torch.Size([1, 8, 74, 128]) - layer.1.v_cache: torch.Size([1, 8, 74, 128]) - layer.2.k_cache: torch.Size([1, 8, 74, 128]) - layer.2.v_cache: torch.Size([1, 8, 74, 128]) - layer.3.k_cache: torch.Size([1, 8, 74, 128]) - layer.3.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.k_cache: torch.Size([1, 8, 74, 128]) - layer.4.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.output: torch.Size([1, 74, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.286s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 74, 128]) - layer.0.v_cache: torch.Size([1, 8, 74, 128]) - layer.1.k_cache: torch.Size([1, 8, 74, 128]) - layer.1.v_cache: torch.Size([1, 8, 74, 128]) - layer.2.k_cache: torch.Size([1, 8, 74, 128]) - layer.2.v_cache: torch.Size([1, 8, 74, 128]) - layer.3.k_cache: torch.Size([1, 8, 74, 128]) - layer.3.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.k_cache: torch.Size([1, 8, 74, 128]) - layer.4.v_cache: torch.Size([1, 8, 74, 128]) - layer.4.output: torch.Size([1, 74, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02555943 33.01170555 - layer.0.v_cache 0.00000028 0.00032435 - layer.1.k_cache 0.00343224 1.87281964 - layer.1.v_cache 0.00000079 0.00113606 - layer.2.k_cache 0.00111603 0.78493623 - layer.2.v_cache 0.00000116 0.00164933 - layer.3.k_cache 0.00133047 0.88077638 - layer.3.v_cache 0.00000218 0.00268155 - layer.4.k_cache 0.00334566 1.75768837 - layer.4.v_cache 0.00000301 0.00423772 - layer.4.output 0.00024685 0.14212211 - ------------------------------------------------------------------------------------- - TOTAL 0.00255562 2.77760312 - (elements=1,060,864) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1060864 -Total Bytes 154616 -BPFP 1.1660 bits/point -EBPFP 2.3319 equivalent bits/point -MSE 2.777603 ----------------------- -------------------------------------------------------- -Time: 0.495s Load: 0.005s, Pack+Encode: 0.204s, Decode+Unpack: 0.286s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 74, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 74, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.7776 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample777-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample777-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample778-layer4-item1.zst (89/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample778-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 61, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.004s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 61, 128) -Output shape: (1, 61, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.0.v_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.1.k_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.1.v_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.2.k_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.2.v_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.3.k_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.3.v_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.4.k_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.4.v_cache: torch.Size([1, 8, 61, 128]) -> torch.Size([1, 1, 61, 1024]) - layer.4.output: torch.Size([1, 61, 4096]) -> torch.Size([1, 1, 61, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,288B, BPFP=0.2930 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 10,256B, BPFP=1.3135 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 6,088B, BPFP=0.7797 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 11,076B, BPFP=1.4185 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 7,424B, BPFP=0.9508 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 11,580B, BPFP=1.4831 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 7,684B, BPFP=0.9841 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 11,484B, BPFP=1.4708 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 5,536B, BPFP=0.7090 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 11,192B, BPFP=1.4334 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 25,780B, BPFP=0.8254 -⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 61, 128]) - layer.0.v_cache: torch.Size([1, 8, 61, 128]) - layer.1.k_cache: torch.Size([1, 8, 61, 128]) - layer.1.v_cache: torch.Size([1, 8, 61, 128]) - layer.2.k_cache: torch.Size([1, 8, 61, 128]) - layer.2.v_cache: torch.Size([1, 8, 61, 128]) - layer.3.k_cache: torch.Size([1, 8, 61, 128]) - layer.3.v_cache: torch.Size([1, 8, 61, 128]) - layer.4.k_cache: torch.Size([1, 8, 61, 128]) - layer.4.v_cache: torch.Size([1, 8, 61, 128]) - layer.4.output: torch.Size([1, 61, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.204s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 61, 128]) - layer.0.v_cache: torch.Size([1, 8, 61, 128]) - layer.1.k_cache: torch.Size([1, 8, 61, 128]) - layer.1.v_cache: torch.Size([1, 8, 61, 128]) - layer.2.k_cache: torch.Size([1, 8, 61, 128]) - layer.2.v_cache: torch.Size([1, 8, 61, 128]) - layer.3.k_cache: torch.Size([1, 8, 61, 128]) - layer.3.v_cache: torch.Size([1, 8, 61, 128]) - layer.4.k_cache: torch.Size([1, 8, 61, 128]) - layer.4.v_cache: torch.Size([1, 8, 61, 128]) - layer.4.output: torch.Size([1, 61, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02755715 22.79067943 - layer.0.v_cache 0.00000026 0.00028992 - layer.1.k_cache 0.00385674 1.51960417 - layer.1.v_cache 0.00000079 0.00115118 - layer.2.k_cache 0.00117960 0.72412460 - layer.2.v_cache 0.00000113 0.00166649 - layer.3.k_cache 0.00144752 0.83093199 - layer.3.v_cache 0.00000212 0.00263009 - layer.4.k_cache 0.00316257 1.57808998 - layer.4.v_cache 0.00000306 0.00425331 - layer.4.output 0.00025867 0.13596622 - ------------------------------------------------------------------------------------- - TOTAL 0.00273183 1.99980615 - (elements=874,496) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 874496 -Total Bytes 110388 -BPFP 1.0098 bits/point -EBPFP 2.0197 equivalent bits/point -MSE 1.999806 ----------------------- -------------------------------------------------------- -Time: 0.363s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.204s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 61, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 61, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9998 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample778-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample778-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample807-layer4-item1.zst (90/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample807-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 59, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 59, 128) -Output shape: (1, 59, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.0.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.1.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.1.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.2.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.2.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.3.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.3.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.4.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.4.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.4.output: torch.Size([1, 59, 4096]) -> torch.Size([1, 1, 59, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,152B, BPFP=0.2850 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 9,952B, BPFP=1.3178 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 5,664B, BPFP=0.7500 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 10,692B, BPFP=1.4158 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 6,924B, BPFP=0.9168 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 10,916B, BPFP=1.4454 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 7,148B, BPFP=0.9465 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 10,984B, BPFP=1.4544 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 5,092B, BPFP=0.6743 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 10,900B, BPFP=1.4433 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 25,076B, BPFP=0.8301 -⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 59, 128]) - layer.0.v_cache: torch.Size([1, 8, 59, 128]) - layer.1.k_cache: torch.Size([1, 8, 59, 128]) - layer.1.v_cache: torch.Size([1, 8, 59, 128]) - layer.2.k_cache: torch.Size([1, 8, 59, 128]) - layer.2.v_cache: torch.Size([1, 8, 59, 128]) - layer.3.k_cache: torch.Size([1, 8, 59, 128]) - layer.3.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.k_cache: torch.Size([1, 8, 59, 128]) - layer.4.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.output: torch.Size([1, 59, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.202s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 59, 128]) - layer.0.v_cache: torch.Size([1, 8, 59, 128]) - layer.1.k_cache: torch.Size([1, 8, 59, 128]) - layer.1.v_cache: torch.Size([1, 8, 59, 128]) - layer.2.k_cache: torch.Size([1, 8, 59, 128]) - layer.2.v_cache: torch.Size([1, 8, 59, 128]) - layer.3.k_cache: torch.Size([1, 8, 59, 128]) - layer.3.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.k_cache: torch.Size([1, 8, 59, 128]) - layer.4.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.output: torch.Size([1, 59, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03100202 26.10237354 - layer.0.v_cache 0.00000029 0.00033417 - layer.1.k_cache 0.00372036 1.70260905 - layer.1.v_cache 0.00000085 0.00124792 - layer.2.k_cache 0.00115281 0.79406965 - layer.2.v_cache 0.00000145 0.00186506 - layer.3.k_cache 0.00140016 0.90778500 - layer.3.v_cache 0.00000257 0.00314998 - layer.4.k_cache 0.00323529 1.79693461 - layer.4.v_cache 0.00000321 0.00481159 - layer.4.output 0.00024217 0.17218721 - ------------------------------------------------------------------------------------- - TOTAL 0.00296341 2.28599496 - (elements=845,824) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 845824 -Total Bytes 105500 -BPFP 0.9978 bits/point -EBPFP 1.9957 equivalent bits/point -MSE 2.285995 ----------------------- -------------------------------------------------------- -Time: 0.362s Load: 0.005s, Pack+Encode: 0.155s, Decode+Unpack: 0.202s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 59, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2860 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample807-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample807-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample855-layer4-item1.zst (91/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample855-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 66, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 66, 128) -Output shape: (1, 66, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.0.v_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.1.k_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.1.v_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.2.k_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.2.v_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.3.k_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.3.v_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.4.k_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.4.v_cache: torch.Size([1, 8, 66, 128]) -> torch.Size([1, 1, 66, 1024]) - layer.4.output: torch.Size([1, 66, 4096]) -> torch.Size([1, 1, 66, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,316B, BPFP=0.2741 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,216B, BPFP=1.5644 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,072B, BPFP=0.8371 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 14,480B, BPFP=1.7140 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 9,272B, BPFP=1.0975 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 15,024B, BPFP=1.7784 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 9,220B, BPFP=1.0914 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 14,608B, BPFP=1.7292 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 5,664B, BPFP=0.6705 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 14,832B, BPFP=1.7557 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 23,820B, BPFP=0.7049 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 66, 128]) - layer.0.v_cache: torch.Size([1, 8, 66, 128]) - layer.1.k_cache: torch.Size([1, 8, 66, 128]) - layer.1.v_cache: torch.Size([1, 8, 66, 128]) - layer.2.k_cache: torch.Size([1, 8, 66, 128]) - layer.2.v_cache: torch.Size([1, 8, 66, 128]) - layer.3.k_cache: torch.Size([1, 8, 66, 128]) - layer.3.v_cache: torch.Size([1, 8, 66, 128]) - layer.4.k_cache: torch.Size([1, 8, 66, 128]) - layer.4.v_cache: torch.Size([1, 8, 66, 128]) - layer.4.output: torch.Size([1, 66, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.291s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 66, 128]) - layer.0.v_cache: torch.Size([1, 8, 66, 128]) - layer.1.k_cache: torch.Size([1, 8, 66, 128]) - layer.1.v_cache: torch.Size([1, 8, 66, 128]) - layer.2.k_cache: torch.Size([1, 8, 66, 128]) - layer.2.v_cache: torch.Size([1, 8, 66, 128]) - layer.3.k_cache: torch.Size([1, 8, 66, 128]) - layer.3.v_cache: torch.Size([1, 8, 66, 128]) - layer.4.k_cache: torch.Size([1, 8, 66, 128]) - layer.4.v_cache: torch.Size([1, 8, 66, 128]) - layer.4.output: torch.Size([1, 66, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02815026 29.75021455 - layer.0.v_cache 0.00000028 0.00031373 - layer.1.k_cache 0.00365239 1.83599171 - layer.1.v_cache 0.00000074 0.00110555 - layer.2.k_cache 0.00118996 0.80249983 - layer.2.v_cache 0.00000101 0.00149319 - layer.3.k_cache 0.00141080 0.90798153 - layer.3.v_cache 0.00000191 0.00245994 - layer.4.k_cache 0.00329622 1.95684676 - layer.4.v_cache 0.00000294 0.00416363 - layer.4.output 0.00019151 0.16610099 - ------------------------------------------------------------------------------------- - TOTAL 0.00274804 2.56624817 - (elements=946,176) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 946176 -Total Bytes 129524 -BPFP 1.0951 bits/point -EBPFP 2.1903 equivalent bits/point -MSE 2.566248 ----------------------- -------------------------------------------------------- -Time: 0.500s Load: 0.005s, Pack+Encode: 0.204s, Decode+Unpack: 0.291s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 66, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 66, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5662 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample855-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample855-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample859-layer4-item1.zst (92/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample859-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 73, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 73, 128) -Output shape: (1, 73, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) -> torch.Size([1, 1, 73, 1024]) - layer.4.output: torch.Size([1, 73, 4096]) -> torch.Size([1, 1, 73, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,408B, BPFP=0.2577 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 13,848B, BPFP=1.4820 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 7,672B, BPFP=0.8211 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 15,200B, BPFP=1.6267 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 10,136B, BPFP=1.0848 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 16,084B, BPFP=1.7213 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 10,400B, BPFP=1.1130 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 15,696B, BPFP=1.6798 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 6,688B, BPFP=0.7158 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 15,728B, BPFP=1.6832 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 32,884B, BPFP=0.8798 -⌛️ [2/4] FRONTEND: Frontend time: 0.204s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.output: torch.Size([1, 73, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.287s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 73, 128]) - layer.0.v_cache: torch.Size([1, 8, 73, 128]) - layer.1.k_cache: torch.Size([1, 8, 73, 128]) - layer.1.v_cache: torch.Size([1, 8, 73, 128]) - layer.2.k_cache: torch.Size([1, 8, 73, 128]) - layer.2.v_cache: torch.Size([1, 8, 73, 128]) - layer.3.k_cache: torch.Size([1, 8, 73, 128]) - layer.3.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.k_cache: torch.Size([1, 8, 73, 128]) - layer.4.v_cache: torch.Size([1, 8, 73, 128]) - layer.4.output: torch.Size([1, 73, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02483096 32.67126967 - layer.0.v_cache 0.00000028 0.00032357 - layer.1.k_cache 0.00355597 1.95727309 - layer.1.v_cache 0.00000080 0.00110745 - layer.2.k_cache 0.00115478 0.76228704 - layer.2.v_cache 0.00000105 0.00154374 - layer.3.k_cache 0.00135922 0.87631769 - layer.3.v_cache 0.00000200 0.00249280 - layer.4.k_cache 0.00333109 1.80706139 - layer.4.v_cache 0.00000279 0.00406897 - layer.4.output 0.00023199 0.13786568 - ------------------------------------------------------------------------------------- - TOTAL 0.00251192 2.75965772 - (elements=1,046,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1046528 -Total Bytes 146744 -BPFP 1.1218 bits/point -EBPFP 2.2435 equivalent bits/point -MSE 2.759658 ----------------------- -------------------------------------------------------- -Time: 0.496s Load: 0.005s, Pack+Encode: 0.204s, Decode+Unpack: 0.287s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 73, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 73, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.7597 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample859-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample859-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample86-layer4-item1.zst (93/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample86-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 121, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 121, 128) -Output shape: (1, 121, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.0.v_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.1.k_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.1.v_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.2.k_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.2.v_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.3.k_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.3.v_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.4.k_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.4.v_cache: torch.Size([1, 8, 121, 128]) -> torch.Size([1, 1, 121, 1024]) - layer.4.output: torch.Size([1, 121, 4096]) -> torch.Size([1, 1, 121, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,292B, BPFP=0.2126 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,456B, BPFP=1.3853 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,388B, BPFP=0.6707 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 23,064B, BPFP=1.4892 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,044B, BPFP=0.8422 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 23,236B, BPFP=1.5003 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 14,044B, BPFP=0.9068 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 22,164B, BPFP=1.4310 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,148B, BPFP=0.6552 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,620B, BPFP=1.5251 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 63,288B, BPFP=1.0216 -⌛️ [2/4] FRONTEND: Frontend time: 0.207s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 121, 128]) - layer.0.v_cache: torch.Size([1, 8, 121, 128]) - layer.1.k_cache: torch.Size([1, 8, 121, 128]) - layer.1.v_cache: torch.Size([1, 8, 121, 128]) - layer.2.k_cache: torch.Size([1, 8, 121, 128]) - layer.2.v_cache: torch.Size([1, 8, 121, 128]) - layer.3.k_cache: torch.Size([1, 8, 121, 128]) - layer.3.v_cache: torch.Size([1, 8, 121, 128]) - layer.4.k_cache: torch.Size([1, 8, 121, 128]) - layer.4.v_cache: torch.Size([1, 8, 121, 128]) - layer.4.output: torch.Size([1, 121, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 121, 128]) - layer.0.v_cache: torch.Size([1, 8, 121, 128]) - layer.1.k_cache: torch.Size([1, 8, 121, 128]) - layer.1.v_cache: torch.Size([1, 8, 121, 128]) - layer.2.k_cache: torch.Size([1, 8, 121, 128]) - layer.2.v_cache: torch.Size([1, 8, 121, 128]) - layer.3.k_cache: torch.Size([1, 8, 121, 128]) - layer.3.v_cache: torch.Size([1, 8, 121, 128]) - layer.4.k_cache: torch.Size([1, 8, 121, 128]) - layer.4.v_cache: torch.Size([1, 8, 121, 128]) - layer.4.output: torch.Size([1, 121, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02495229 23.01708581 - layer.0.v_cache 0.00000028 0.00030693 - layer.1.k_cache 0.00332455 1.43627425 - layer.1.v_cache 0.00000080 0.00107568 - layer.2.k_cache 0.00116246 0.69485070 - layer.2.v_cache 0.00000109 0.00148402 - layer.3.k_cache 0.00134746 0.77986227 - layer.3.v_cache 0.00000209 0.00253847 - layer.4.k_cache 0.00345892 1.51722995 - layer.4.v_cache 0.00000309 0.00423552 - layer.4.output 0.00018371 0.10884260 - ------------------------------------------------------------------------------------- - TOTAL 0.00249913 1.99216529 - (elements=1,734,656) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1734656 -Total Bytes 227744 -BPFP 1.0503 bits/point -EBPFP 2.1006 equivalent bits/point -MSE 1.992165 ----------------------- -------------------------------------------------------- -Time: 0.503s Load: 0.007s, Pack+Encode: 0.207s, Decode+Unpack: 0.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 121, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 121, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9922 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample86-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample86-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample91-layer4-item1.zst (94/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample91-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 113, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 113, 128) -Output shape: (1, 113, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.0.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.1.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.1.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.2.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.2.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.3.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.3.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.4.k_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.4.v_cache: torch.Size([1, 8, 113, 128]) -> torch.Size([1, 1, 113, 1024]) - layer.4.output: torch.Size([1, 113, 4096]) -> torch.Size([1, 1, 113, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,348B, BPFP=0.2315 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,208B, BPFP=1.4663 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,720B, BPFP=0.7412 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,564B, BPFP=1.5600 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,596B, BPFP=0.8709 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,496B, BPFP=1.5553 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,624B, BPFP=0.9419 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,880B, BPFP=1.5127 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,824B, BPFP=0.8175 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 22,836B, BPFP=1.5788 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 53,772B, BPFP=0.9294 -⌛️ [2/4] FRONTEND: Frontend time: 0.206s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 113, 128]) - layer.0.v_cache: torch.Size([1, 8, 113, 128]) - layer.1.k_cache: torch.Size([1, 8, 113, 128]) - layer.1.v_cache: torch.Size([1, 8, 113, 128]) - layer.2.k_cache: torch.Size([1, 8, 113, 128]) - layer.2.v_cache: torch.Size([1, 8, 113, 128]) - layer.3.k_cache: torch.Size([1, 8, 113, 128]) - layer.3.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.k_cache: torch.Size([1, 8, 113, 128]) - layer.4.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.output: torch.Size([1, 113, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 113, 128]) - layer.0.v_cache: torch.Size([1, 8, 113, 128]) - layer.1.k_cache: torch.Size([1, 8, 113, 128]) - layer.1.v_cache: torch.Size([1, 8, 113, 128]) - layer.2.k_cache: torch.Size([1, 8, 113, 128]) - layer.2.v_cache: torch.Size([1, 8, 113, 128]) - layer.3.k_cache: torch.Size([1, 8, 113, 128]) - layer.3.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.k_cache: torch.Size([1, 8, 113, 128]) - layer.4.v_cache: torch.Size([1, 8, 113, 128]) - layer.4.output: torch.Size([1, 113, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02530831 24.91992404 - layer.0.v_cache 0.00000029 0.00030710 - layer.1.k_cache 0.00330439 1.60623277 - layer.1.v_cache 0.00000076 0.00105881 - layer.2.k_cache 0.00114562 0.72289107 - layer.2.v_cache 0.00000109 0.00147682 - layer.3.k_cache 0.00136171 0.84187891 - layer.3.v_cache 0.00000227 0.00268610 - layer.4.k_cache 0.00342222 1.58541019 - layer.4.v_cache 0.00000313 0.00420005 - layer.4.output 0.00018801 0.11914904 - ------------------------------------------------------------------------------------- - TOTAL 0.00252156 2.15447586 - (elements=1,619,968) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1619968 -Total Bytes 216868 -BPFP 1.0710 bits/point -EBPFP 2.1419 equivalent bits/point -MSE 2.154476 ----------------------- -------------------------------------------------------- -Time: 0.501s Load: 0.008s, Pack+Encode: 0.206s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 113, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 113, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1545 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample91-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample91-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample92-layer4-item1.zst (95/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample92-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 110, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 110, 128) -Output shape: (1, 110, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.0.v_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.1.k_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.1.v_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.2.k_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.2.v_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.3.k_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.3.v_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.4.k_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.4.v_cache: torch.Size([1, 8, 110, 128]) -> torch.Size([1, 1, 110, 1024]) - layer.4.output: torch.Size([1, 110, 4096]) -> torch.Size([1, 1, 110, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,188B, BPFP=0.2264 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 21,108B, BPFP=1.4991 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,104B, BPFP=0.7176 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,308B, BPFP=1.5844 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,424B, BPFP=0.9534 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,676B, BPFP=1.6105 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,616B, BPFP=0.9670 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,940B, BPFP=1.5582 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,332B, BPFP=0.8048 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,068B, BPFP=1.6384 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 53,884B, BPFP=0.9567 -⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 110, 128]) - layer.0.v_cache: torch.Size([1, 8, 110, 128]) - layer.1.k_cache: torch.Size([1, 8, 110, 128]) - layer.1.v_cache: torch.Size([1, 8, 110, 128]) - layer.2.k_cache: torch.Size([1, 8, 110, 128]) - layer.2.v_cache: torch.Size([1, 8, 110, 128]) - layer.3.k_cache: torch.Size([1, 8, 110, 128]) - layer.3.v_cache: torch.Size([1, 8, 110, 128]) - layer.4.k_cache: torch.Size([1, 8, 110, 128]) - layer.4.v_cache: torch.Size([1, 8, 110, 128]) - layer.4.output: torch.Size([1, 110, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.283s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 110, 128]) - layer.0.v_cache: torch.Size([1, 8, 110, 128]) - layer.1.k_cache: torch.Size([1, 8, 110, 128]) - layer.1.v_cache: torch.Size([1, 8, 110, 128]) - layer.2.k_cache: torch.Size([1, 8, 110, 128]) - layer.2.v_cache: torch.Size([1, 8, 110, 128]) - layer.3.k_cache: torch.Size([1, 8, 110, 128]) - layer.3.v_cache: torch.Size([1, 8, 110, 128]) - layer.4.k_cache: torch.Size([1, 8, 110, 128]) - layer.4.v_cache: torch.Size([1, 8, 110, 128]) - layer.4.output: torch.Size([1, 110, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02547969 25.82180398 - layer.0.v_cache 0.00000028 0.00030062 - layer.1.k_cache 0.00323236 1.55474978 - layer.1.v_cache 0.00000082 0.00105486 - layer.2.k_cache 0.00114874 0.73921752 - layer.2.v_cache 0.00000131 0.00156044 - layer.3.k_cache 0.00138397 0.82285656 - layer.3.v_cache 0.00000212 0.00257366 - layer.4.k_cache 0.00346958 1.70442963 - layer.4.v_cache 0.00000311 0.00420452 - layer.4.output 0.00021756 0.12129381 - ------------------------------------------------------------------------------------- - TOTAL 0.00254230 2.22413763 - (elements=1,576,960) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1576960 -Total Bytes 216648 -BPFP 1.0991 bits/point -EBPFP 2.1981 equivalent bits/point -MSE 2.224138 ----------------------- -------------------------------------------------------- -Time: 0.496s Load: 0.006s, Pack+Encode: 0.208s, Decode+Unpack: 0.283s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 110, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 110, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2241 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample92-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample92-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample925-layer4-item1.zst (96/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample925-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 65, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 65, 128) -Output shape: (1, 65, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.0.v_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.1.k_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.1.v_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.2.k_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.2.v_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.3.k_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.3.v_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.4.k_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.4.v_cache: torch.Size([1, 8, 65, 128]) -> torch.Size([1, 1, 65, 1024]) - layer.4.output: torch.Size([1, 65, 4096]) -> torch.Size([1, 1, 65, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,236B, BPFP=0.2687 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 12,592B, BPFP=1.5135 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 6,492B, BPFP=0.7803 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 13,744B, BPFP=1.6519 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 8,352B, BPFP=1.0038 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 13,784B, BPFP=1.6567 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 8,732B, BPFP=1.0495 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 14,436B, BPFP=1.7351 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 5,476B, BPFP=0.6582 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 14,392B, BPFP=1.7298 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 25,304B, BPFP=0.7603 -⌛️ [2/4] FRONTEND: Frontend time: 0.198s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 65, 128]) - layer.0.v_cache: torch.Size([1, 8, 65, 128]) - layer.1.k_cache: torch.Size([1, 8, 65, 128]) - layer.1.v_cache: torch.Size([1, 8, 65, 128]) - layer.2.k_cache: torch.Size([1, 8, 65, 128]) - layer.2.v_cache: torch.Size([1, 8, 65, 128]) - layer.3.k_cache: torch.Size([1, 8, 65, 128]) - layer.3.v_cache: torch.Size([1, 8, 65, 128]) - layer.4.k_cache: torch.Size([1, 8, 65, 128]) - layer.4.v_cache: torch.Size([1, 8, 65, 128]) - layer.4.output: torch.Size([1, 65, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.324s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 65, 128]) - layer.0.v_cache: torch.Size([1, 8, 65, 128]) - layer.1.k_cache: torch.Size([1, 8, 65, 128]) - layer.1.v_cache: torch.Size([1, 8, 65, 128]) - layer.2.k_cache: torch.Size([1, 8, 65, 128]) - layer.2.v_cache: torch.Size([1, 8, 65, 128]) - layer.3.k_cache: torch.Size([1, 8, 65, 128]) - layer.3.v_cache: torch.Size([1, 8, 65, 128]) - layer.4.k_cache: torch.Size([1, 8, 65, 128]) - layer.4.v_cache: torch.Size([1, 8, 65, 128]) - layer.4.output: torch.Size([1, 65, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02613364 29.34725436 - layer.0.v_cache 0.00000028 0.00032826 - layer.1.k_cache 0.00351263 1.83675396 - layer.1.v_cache 0.00000086 0.00111491 - layer.2.k_cache 0.00113140 0.78985842 - layer.2.v_cache 0.00000103 0.00154457 - layer.3.k_cache 0.00134002 0.85776308 - layer.3.v_cache 0.00000212 0.00265086 - layer.4.k_cache 0.00331905 1.83258421 - layer.4.v_cache 0.00000293 0.00426551 - layer.4.output 0.00019456 0.14365198 - ------------------------------------------------------------------------------------- - TOTAL 0.00258730 2.51776615 - (elements=931,840) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 931840 -Total Bytes 125540 -BPFP 1.0778 bits/point -EBPFP 2.1556 equivalent bits/point -MSE 2.517766 ----------------------- -------------------------------------------------------- -Time: 0.527s Load: 0.005s, Pack+Encode: 0.198s, Decode+Unpack: 0.324s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 65, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 65, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.5178 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample925-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample925-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample95-layer4-item1.zst (97/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample95-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 116, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 116, 128) -Output shape: (1, 116, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.0.v_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.1.k_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.1.v_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.2.k_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.2.v_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.3.k_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.3.v_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.4.k_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.4.v_cache: torch.Size([1, 8, 116, 128]) -> torch.Size([1, 1, 116, 1024]) - layer.4.output: torch.Size([1, 116, 4096]) -> torch.Size([1, 1, 116, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,340B, BPFP=0.2249 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,900B, BPFP=1.4076 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,124B, BPFP=0.6818 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,512B, BPFP=1.5162 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,544B, BPFP=0.8448 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,652B, BPFP=1.5256 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,440B, BPFP=0.9052 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,808B, BPFP=1.4688 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,044B, BPFP=0.7438 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,464B, BPFP=1.5803 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 67,216B, BPFP=1.1317 -⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 116, 128]) - layer.0.v_cache: torch.Size([1, 8, 116, 128]) - layer.1.k_cache: torch.Size([1, 8, 116, 128]) - layer.1.v_cache: torch.Size([1, 8, 116, 128]) - layer.2.k_cache: torch.Size([1, 8, 116, 128]) - layer.2.v_cache: torch.Size([1, 8, 116, 128]) - layer.3.k_cache: torch.Size([1, 8, 116, 128]) - layer.3.v_cache: torch.Size([1, 8, 116, 128]) - layer.4.k_cache: torch.Size([1, 8, 116, 128]) - layer.4.v_cache: torch.Size([1, 8, 116, 128]) - layer.4.output: torch.Size([1, 116, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 116, 128]) - layer.0.v_cache: torch.Size([1, 8, 116, 128]) - layer.1.k_cache: torch.Size([1, 8, 116, 128]) - layer.1.v_cache: torch.Size([1, 8, 116, 128]) - layer.2.k_cache: torch.Size([1, 8, 116, 128]) - layer.2.v_cache: torch.Size([1, 8, 116, 128]) - layer.3.k_cache: torch.Size([1, 8, 116, 128]) - layer.3.v_cache: torch.Size([1, 8, 116, 128]) - layer.4.k_cache: torch.Size([1, 8, 116, 128]) - layer.4.v_cache: torch.Size([1, 8, 116, 128]) - layer.4.output: torch.Size([1, 116, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02494503 22.07805344 - layer.0.v_cache 0.00000027 0.00030888 - layer.1.k_cache 0.00326550 1.55408767 - layer.1.v_cache 0.00000084 0.00112107 - layer.2.k_cache 0.00114506 0.73430857 - layer.2.v_cache 0.00000107 0.00155233 - layer.3.k_cache 0.00140586 0.83080410 - layer.3.v_cache 0.00000237 0.00270434 - layer.4.k_cache 0.00329219 1.63468920 - layer.4.v_cache 0.00000320 0.00436717 - layer.4.output 0.00021339 0.13784491 - ------------------------------------------------------------------------------------- - TOTAL 0.00249392 1.95666974 - (elements=1,662,976) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1662976 -Total Bytes 229044 -BPFP 1.1019 bits/point -EBPFP 2.2037 equivalent bits/point -MSE 1.956670 ----------------------- -------------------------------------------------------- -Time: 0.514s Load: 0.008s, Pack+Encode: 0.219s, Decode+Unpack: 0.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 116, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 116, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.9567 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample95-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample95-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample967-layer4-item1.zst (98/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample967-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 59, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.004s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 59, 128) -Output shape: (1, 59, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.0.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.1.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.1.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.2.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.2.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.3.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.3.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.4.k_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.4.v_cache: torch.Size([1, 8, 59, 128]) -> torch.Size([1, 1, 59, 1024]) - layer.4.output: torch.Size([1, 59, 4096]) -> torch.Size([1, 1, 59, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,120B, BPFP=0.2807 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 9,864B, BPFP=1.3061 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 5,484B, BPFP=0.7262 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 10,660B, BPFP=1.4115 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 6,700B, BPFP=0.8872 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 11,100B, BPFP=1.4698 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 6,904B, BPFP=0.9142 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 10,656B, BPFP=1.4110 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 4,928B, BPFP=0.6525 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 10,756B, BPFP=1.4243 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 24,232B, BPFP=0.8022 -⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 59, 128]) - layer.0.v_cache: torch.Size([1, 8, 59, 128]) - layer.1.k_cache: torch.Size([1, 8, 59, 128]) - layer.1.v_cache: torch.Size([1, 8, 59, 128]) - layer.2.k_cache: torch.Size([1, 8, 59, 128]) - layer.2.v_cache: torch.Size([1, 8, 59, 128]) - layer.3.k_cache: torch.Size([1, 8, 59, 128]) - layer.3.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.k_cache: torch.Size([1, 8, 59, 128]) - layer.4.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.output: torch.Size([1, 59, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.204s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 59, 128]) - layer.0.v_cache: torch.Size([1, 8, 59, 128]) - layer.1.k_cache: torch.Size([1, 8, 59, 128]) - layer.1.v_cache: torch.Size([1, 8, 59, 128]) - layer.2.k_cache: torch.Size([1, 8, 59, 128]) - layer.2.v_cache: torch.Size([1, 8, 59, 128]) - layer.3.k_cache: torch.Size([1, 8, 59, 128]) - layer.3.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.k_cache: torch.Size([1, 8, 59, 128]) - layer.4.v_cache: torch.Size([1, 8, 59, 128]) - layer.4.output: torch.Size([1, 59, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02763205 25.03834249 - layer.0.v_cache 0.00000028 0.00031663 - layer.1.k_cache 0.00373147 1.77873708 - layer.1.v_cache 0.00000088 0.00109553 - layer.2.k_cache 0.00131257 0.76702596 - layer.2.v_cache 0.00000108 0.00152047 - layer.3.k_cache 0.00142120 0.87736460 - layer.3.v_cache 0.00000202 0.00239246 - layer.4.k_cache 0.00321287 1.87068926 - layer.4.v_cache 0.00000283 0.00391084 - layer.4.output 0.00027608 0.14361225 - ------------------------------------------------------------------------------------- - TOTAL 0.00274440 2.20827459 - (elements=845,824) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 845824 -Total Bytes 103404 -BPFP 0.9780 bits/point -EBPFP 1.9560 equivalent bits/point -MSE 2.208275 ----------------------- -------------------------------------------------------- -Time: 0.364s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.204s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 59, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 59, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.2083 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample967-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample967-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample969-layer4-item1.zst (99/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample969-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 60, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.004s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 60, 128) -Output shape: (1, 60, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.0.v_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.1.k_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.1.v_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.2.k_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.2.v_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.3.k_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.3.v_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.4.k_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.4.v_cache: torch.Size([1, 8, 60, 128]) -> torch.Size([1, 1, 60, 1024]) - layer.4.output: torch.Size([1, 60, 4096]) -> torch.Size([1, 1, 60, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,168B, BPFP=0.2823 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 9,980B, BPFP=1.2995 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 5,580B, BPFP=0.7266 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 10,864B, BPFP=1.4146 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 6,864B, BPFP=0.8938 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 11,204B, BPFP=1.4589 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 6,944B, BPFP=0.9042 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 10,848B, BPFP=1.4125 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 5,032B, BPFP=0.6552 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 11,100B, BPFP=1.4453 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 25,136B, BPFP=0.8182 -⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 60, 128]) - layer.0.v_cache: torch.Size([1, 8, 60, 128]) - layer.1.k_cache: torch.Size([1, 8, 60, 128]) - layer.1.v_cache: torch.Size([1, 8, 60, 128]) - layer.2.k_cache: torch.Size([1, 8, 60, 128]) - layer.2.v_cache: torch.Size([1, 8, 60, 128]) - layer.3.k_cache: torch.Size([1, 8, 60, 128]) - layer.3.v_cache: torch.Size([1, 8, 60, 128]) - layer.4.k_cache: torch.Size([1, 8, 60, 128]) - layer.4.v_cache: torch.Size([1, 8, 60, 128]) - layer.4.output: torch.Size([1, 60, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.201s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 60, 128]) - layer.0.v_cache: torch.Size([1, 8, 60, 128]) - layer.1.k_cache: torch.Size([1, 8, 60, 128]) - layer.1.v_cache: torch.Size([1, 8, 60, 128]) - layer.2.k_cache: torch.Size([1, 8, 60, 128]) - layer.2.v_cache: torch.Size([1, 8, 60, 128]) - layer.3.k_cache: torch.Size([1, 8, 60, 128]) - layer.3.v_cache: torch.Size([1, 8, 60, 128]) - layer.4.k_cache: torch.Size([1, 8, 60, 128]) - layer.4.v_cache: torch.Size([1, 8, 60, 128]) - layer.4.output: torch.Size([1, 60, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02844323 25.09667969 - layer.0.v_cache 0.00000027 0.00031170 - layer.1.k_cache 0.00360089 1.74303907 - layer.1.v_cache 0.00000087 0.00115058 - layer.2.k_cache 0.00116909 0.73845177 - layer.2.v_cache 0.00000119 0.00167237 - layer.3.k_cache 0.00139808 0.86212292 - layer.3.v_cache 0.00000212 0.00265158 - layer.4.k_cache 0.00324009 1.79383367 - layer.4.v_cache 0.00000310 0.00418026 - layer.4.output 0.00020032 0.13090754 - ------------------------------------------------------------------------------------- - TOTAL 0.00276144 2.19769455 - (elements=860,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 860160 -Total Bytes 105720 -BPFP 0.9833 bits/point -EBPFP 1.9665 equivalent bits/point -MSE 2.197695 ----------------------- -------------------------------------------------------- -Time: 0.360s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.201s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 60, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 60, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1977 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample969-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample969-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample97-layer4-item1.zst (100/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample97-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 114, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 114, 128) -Output shape: (1, 114, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.0.v_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.1.k_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.1.v_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.2.k_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.2.v_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.3.k_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.3.v_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.4.k_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.4.v_cache: torch.Size([1, 8, 114, 128]) -> torch.Size([1, 1, 114, 1024]) - layer.4.output: torch.Size([1, 114, 4096]) -> torch.Size([1, 1, 114, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,300B, BPFP=0.2262 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 20,772B, BPFP=1.4235 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,348B, BPFP=0.7092 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 22,540B, BPFP=1.5447 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,564B, BPFP=0.8610 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 22,836B, BPFP=1.5650 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 13,536B, BPFP=0.9276 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 21,916B, BPFP=1.5019 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,616B, BPFP=0.7961 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 23,268B, BPFP=1.5946 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 57,468B, BPFP=0.9846 -⌛️ [2/4] FRONTEND: Frontend time: 0.207s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 114, 128]) - layer.0.v_cache: torch.Size([1, 8, 114, 128]) - layer.1.k_cache: torch.Size([1, 8, 114, 128]) - layer.1.v_cache: torch.Size([1, 8, 114, 128]) - layer.2.k_cache: torch.Size([1, 8, 114, 128]) - layer.2.v_cache: torch.Size([1, 8, 114, 128]) - layer.3.k_cache: torch.Size([1, 8, 114, 128]) - layer.3.v_cache: torch.Size([1, 8, 114, 128]) - layer.4.k_cache: torch.Size([1, 8, 114, 128]) - layer.4.v_cache: torch.Size([1, 8, 114, 128]) - layer.4.output: torch.Size([1, 114, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.291s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 114, 128]) - layer.0.v_cache: torch.Size([1, 8, 114, 128]) - layer.1.k_cache: torch.Size([1, 8, 114, 128]) - layer.1.v_cache: torch.Size([1, 8, 114, 128]) - layer.2.k_cache: torch.Size([1, 8, 114, 128]) - layer.2.v_cache: torch.Size([1, 8, 114, 128]) - layer.3.k_cache: torch.Size([1, 8, 114, 128]) - layer.3.v_cache: torch.Size([1, 8, 114, 128]) - layer.4.k_cache: torch.Size([1, 8, 114, 128]) - layer.4.v_cache: torch.Size([1, 8, 114, 128]) - layer.4.output: torch.Size([1, 114, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02435601 25.47337582 - layer.0.v_cache 0.00000028 0.00031229 - layer.1.k_cache 0.00337316 1.52528247 - layer.1.v_cache 0.00000076 0.00111059 - layer.2.k_cache 0.00114524 0.75907898 - layer.2.v_cache 0.00000110 0.00161723 - layer.3.k_cache 0.00139029 0.84935794 - layer.3.v_cache 0.00000222 0.00265852 - layer.4.k_cache 0.00339712 1.68694948 - layer.4.v_cache 0.00000302 0.00436675 - layer.4.output 0.00021100 0.11975201 - ------------------------------------------------------------------------------------- - TOTAL 0.00246523 2.19879415 - (elements=1,634,304) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1634304 -Total Bytes 220164 -BPFP 1.0777 bits/point -EBPFP 2.1554 equivalent bits/point -MSE 2.198794 ----------------------- -------------------------------------------------------- -Time: 0.506s Load: 0.007s, Pack+Encode: 0.207s, Decode+Unpack: 0.291s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 114, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 114, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 2.1988 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_gsm8k/sample97-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample97-layer4-item1.zst ------------------------- ---------------------------- -TOTAL PROCESSING SUMMARY ------------------------- ---------------------------- -Total files 100 -Avg BPFP 1.1466 bits/point -Avg EBPFP 2.2932 equivalent bits/point -Avg MSE 2.248783 -Avg Time 0.656s ------------------------- ---------------------------- +version https://git-lfs.github.com/spec/v1 +oid sha256:e44706655ccdb75a8087260f55bba7d39cd8dc06c2ea0341159999368c5ace46 +size 1118953 diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/dtufc_hyperprior-featurecoding_qwen_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/dtufc_hyperprior-featurecoding_qwen_individual.log index e23ac7f69274cd6bc0185374a945459ae9de462f..3fa591f8634bf98e5a01dfa076e9c452ec496218 100644 --- a/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/dtufc_hyperprior-featurecoding_qwen_individual.log +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/dtufc_hyperprior-featurecoding_qwen_individual.log @@ -1,16958 +1,3 @@ -Experiment: dtufc_hyperprior-featurecoding_qwen_individual -Log file: output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/dtufc_hyperprior-featurecoding_qwen_individual.log -DTUFCCodecConfig: - arch: hyperprior-featurecoding - handler: qwen - checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar - transform_type: kmeans - transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json - bit_depth: 8 - device: cuda:0 -Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar -Checkpoint epoch: 559 -Loaded hyperprior-featurecoding (1-channel) on cuda:0 -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json -Loaded per-key mappings: model=qwen - Keys: ['layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache', 'layer.4.output'] ----------------- ------------------------------------------------------------------------------------------------------------------------------ -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding -Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar -Transform type kmeans -Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json -Input ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag -Output output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag ----------------- ------------------------------------------------------------------------------------------------------------------------------ -Files found: 100 ----------------------------------------------------------------------- - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample0-layer4-item1.zst (1/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample0-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 371, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.026s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 371, 128) -Output shape: (1, 371, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.0.v_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.1.k_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.1.v_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.2.k_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.2.v_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.3.k_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.3.v_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.4.k_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.4.v_cache: torch.Size([1, 8, 371, 128]) -> torch.Size([1, 1, 371, 1024]) - layer.4.output: torch.Size([1, 371, 4096]) -> torch.Size([1, 1, 371, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,272B, BPFP=0.1952 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,216B, BPFP=1.2470 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,004B, BPFP=0.5897 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,944B, BPFP=1.3465 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,152B, BPFP=0.7823 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,144B, BPFP=1.3718 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,728B, BPFP=0.8366 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,456B, BPFP=1.3363 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,660B, BPFP=0.5825 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,064B, BPFP=1.3912 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,884B, BPFP=0.7680 -⌛️ [2/4] FRONTEND: Frontend time: 0.655s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 371, 128]) - layer.0.v_cache: torch.Size([1, 8, 371, 128]) - layer.1.k_cache: torch.Size([1, 8, 371, 128]) - layer.1.v_cache: torch.Size([1, 8, 371, 128]) - layer.2.k_cache: torch.Size([1, 8, 371, 128]) - layer.2.v_cache: torch.Size([1, 8, 371, 128]) - layer.3.k_cache: torch.Size([1, 8, 371, 128]) - layer.3.v_cache: torch.Size([1, 8, 371, 128]) - layer.4.k_cache: torch.Size([1, 8, 371, 128]) - layer.4.v_cache: torch.Size([1, 8, 371, 128]) - layer.4.output: torch.Size([1, 371, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.812s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 371, 128]) - layer.0.v_cache: torch.Size([1, 8, 371, 128]) - layer.1.k_cache: torch.Size([1, 8, 371, 128]) - layer.1.v_cache: torch.Size([1, 8, 371, 128]) - layer.2.k_cache: torch.Size([1, 8, 371, 128]) - layer.2.v_cache: torch.Size([1, 8, 371, 128]) - layer.3.k_cache: torch.Size([1, 8, 371, 128]) - layer.3.v_cache: torch.Size([1, 8, 371, 128]) - layer.4.k_cache: torch.Size([1, 8, 371, 128]) - layer.4.v_cache: torch.Size([1, 8, 371, 128]) - layer.4.output: torch.Size([1, 371, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02382561 17.84119014 - layer.0.v_cache 0.00000026 0.00029091 - layer.1.k_cache 0.00289382 1.20034650 - layer.1.v_cache 0.00000075 0.00107632 - layer.2.k_cache 0.00115144 0.68573146 - layer.2.v_cache 0.00000114 0.00160092 - layer.3.k_cache 0.00133317 0.76193953 - layer.3.v_cache 0.00000213 0.00260576 - layer.4.k_cache 0.00354181 1.40586767 - layer.4.v_cache 0.00000324 0.00434668 - layer.4.output 0.00015639 0.09198427 - ------------------------------------------------------------------------------------- - TOTAL 0.00238421 1.59092378 - (elements=5,318,656) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5318656 -Total Bytes 605524 -BPFP 0.9108 bits/point -EBPFP 1.8216 equivalent bits/point -MSE 1.590924 ----------------------- -------------------------------------------------------- -Time: 1.493s Load: 0.026s, Pack+Encode: 0.655s, Decode+Unpack: 0.812s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 371, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 371, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5909 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample0-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample0-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample1-layer4-item1.zst (2/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample1-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 344, 128) -Output shape: (1, 344, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.output: torch.Size([1, 344, 4096]) -> torch.Size([1, 1, 344, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,988B, BPFP=0.2041 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,604B, BPFP=1.3537 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,140B, BPFP=0.6391 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,272B, BPFP=1.4597 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,476B, BPFP=0.8284 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,928B, BPFP=1.4973 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,540B, BPFP=0.8980 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,820B, BPFP=1.4494 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,092B, BPFP=0.6153 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,308B, BPFP=1.5059 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,496B, BPFP=0.8261 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.693s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02409526 18.95949679 - layer.0.v_cache 0.00000026 0.00029979 - layer.1.k_cache 0.00287302 1.17268860 - layer.1.v_cache 0.00000078 0.00112214 - layer.2.k_cache 0.00123767 0.68556129 - layer.2.v_cache 0.00000117 0.00172167 - layer.3.k_cache 0.00130179 0.74808396 - layer.3.v_cache 0.00000226 0.00278766 - layer.4.k_cache 0.00355144 1.38210217 - layer.4.v_cache 0.00000325 0.00459084 - layer.4.output 0.00015063 0.09236518 - ------------------------------------------------------------------------------------- - TOTAL 0.00240496 1.66627969 - (elements=4,931,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4931584 -Total Bytes 605664 -BPFP 0.9825 bits/point -EBPFP 1.9650 equivalent bits/point -MSE 1.666280 ----------------------- -------------------------------------------------------- -Time: 1.120s Load: 0.019s, Pack+Encode: 0.408s, Decode+Unpack: 0.693s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6663 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample1-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample1-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample10-layer4-item1.zst (3/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample10-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 377, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.021s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 377, 128) -Output shape: (1, 377, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.0.v_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.1.k_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.1.v_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.2.k_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.2.v_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.3.k_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.3.v_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.4.k_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.4.v_cache: torch.Size([1, 8, 377, 128]) -> torch.Size([1, 1, 377, 1024]) - layer.4.output: torch.Size([1, 377, 4096]) -> torch.Size([1, 1, 377, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,488B, BPFP=0.1966 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 60,348B, BPFP=1.2506 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,824B, BPFP=0.5973 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 66,032B, BPFP=1.3684 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 38,548B, BPFP=0.7988 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 66,516B, BPFP=1.3784 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,380B, BPFP=0.8368 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 64,320B, BPFP=1.3329 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,152B, BPFP=0.5834 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 67,692B, BPFP=1.4028 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,264B, BPFP=0.7526 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 377, 128]) - layer.0.v_cache: torch.Size([1, 8, 377, 128]) - layer.1.k_cache: torch.Size([1, 8, 377, 128]) - layer.1.v_cache: torch.Size([1, 8, 377, 128]) - layer.2.k_cache: torch.Size([1, 8, 377, 128]) - layer.2.v_cache: torch.Size([1, 8, 377, 128]) - layer.3.k_cache: torch.Size([1, 8, 377, 128]) - layer.3.v_cache: torch.Size([1, 8, 377, 128]) - layer.4.k_cache: torch.Size([1, 8, 377, 128]) - layer.4.v_cache: torch.Size([1, 8, 377, 128]) - layer.4.output: torch.Size([1, 377, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.685s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 377, 128]) - layer.0.v_cache: torch.Size([1, 8, 377, 128]) - layer.1.k_cache: torch.Size([1, 8, 377, 128]) - layer.1.v_cache: torch.Size([1, 8, 377, 128]) - layer.2.k_cache: torch.Size([1, 8, 377, 128]) - layer.2.v_cache: torch.Size([1, 8, 377, 128]) - layer.3.k_cache: torch.Size([1, 8, 377, 128]) - layer.3.v_cache: torch.Size([1, 8, 377, 128]) - layer.4.k_cache: torch.Size([1, 8, 377, 128]) - layer.4.v_cache: torch.Size([1, 8, 377, 128]) - layer.4.output: torch.Size([1, 377, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02400962 18.36692717 - layer.0.v_cache 0.00000027 0.00030414 - layer.1.k_cache 0.00291640 1.19188304 - layer.1.v_cache 0.00000080 0.00115110 - layer.2.k_cache 0.00116459 0.68230238 - layer.2.v_cache 0.00000114 0.00167149 - layer.3.k_cache 0.00131780 0.73442179 - layer.3.v_cache 0.00000210 0.00266415 - layer.4.k_cache 0.00355012 1.34710928 - layer.4.v_cache 0.00000332 0.00466935 - layer.4.output 0.00014433 0.08735698 - ------------------------------------------------------------------------------------- - TOTAL 0.00239596 1.62018084 - (elements=5,404,672) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5404672 -Total Bytes 615564 -BPFP 0.9112 bits/point -EBPFP 1.8223 equivalent bits/point -MSE 1.620181 ----------------------- -------------------------------------------------------- -Time: 1.111s Load: 0.021s, Pack+Encode: 0.406s, Decode+Unpack: 0.685s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 377, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 377, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6202 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample10-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample10-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample100-layer4-item1.zst (4/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample100-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 315, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 315, 128) -Output shape: (1, 315, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.0.v_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.1.k_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.1.v_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.2.k_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.2.v_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.3.k_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.3.v_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.4.k_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.4.v_cache: torch.Size([1, 8, 315, 128]) -> torch.Size([1, 1, 315, 1024]) - layer.4.output: torch.Size([1, 315, 4096]) -> torch.Size([1, 1, 315, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,172B, BPFP=0.2027 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,708B, BPFP=1.2576 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,012B, BPFP=0.5955 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 54,664B, BPFP=1.3558 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,912B, BPFP=0.7915 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,584B, BPFP=1.3786 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,784B, BPFP=0.8379 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,960B, BPFP=1.3383 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,788B, BPFP=0.5900 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,456B, BPFP=1.4002 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 123,516B, BPFP=0.7658 -⌛️ [2/4] FRONTEND: Frontend time: 0.415s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 315, 128]) - layer.0.v_cache: torch.Size([1, 8, 315, 128]) - layer.1.k_cache: torch.Size([1, 8, 315, 128]) - layer.1.v_cache: torch.Size([1, 8, 315, 128]) - layer.2.k_cache: torch.Size([1, 8, 315, 128]) - layer.2.v_cache: torch.Size([1, 8, 315, 128]) - layer.3.k_cache: torch.Size([1, 8, 315, 128]) - layer.3.v_cache: torch.Size([1, 8, 315, 128]) - layer.4.k_cache: torch.Size([1, 8, 315, 128]) - layer.4.v_cache: torch.Size([1, 8, 315, 128]) - layer.4.output: torch.Size([1, 315, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.590s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 315, 128]) - layer.0.v_cache: torch.Size([1, 8, 315, 128]) - layer.1.k_cache: torch.Size([1, 8, 315, 128]) - layer.1.v_cache: torch.Size([1, 8, 315, 128]) - layer.2.k_cache: torch.Size([1, 8, 315, 128]) - layer.2.v_cache: torch.Size([1, 8, 315, 128]) - layer.3.k_cache: torch.Size([1, 8, 315, 128]) - layer.3.v_cache: torch.Size([1, 8, 315, 128]) - layer.4.k_cache: torch.Size([1, 8, 315, 128]) - layer.4.v_cache: torch.Size([1, 8, 315, 128]) - layer.4.output: torch.Size([1, 315, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02468849 18.57841487 - layer.0.v_cache 0.00000027 0.00030201 - layer.1.k_cache 0.00284817 1.14554511 - layer.1.v_cache 0.00000082 0.00113659 - layer.2.k_cache 0.00122538 0.68706287 - layer.2.v_cache 0.00000116 0.00166647 - layer.3.k_cache 0.00131111 0.75621004 - layer.3.v_cache 0.00000216 0.00275023 - layer.4.k_cache 0.00356899 1.38112357 - layer.4.v_cache 0.00000330 0.00460428 - layer.4.output 0.00013845 0.08972884 - ------------------------------------------------------------------------------------- - TOTAL 0.00244312 1.63698082 - (elements=4,515,840) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4515840 -Total Bytes 516556 -BPFP 0.9151 bits/point -EBPFP 1.8302 equivalent bits/point -MSE 1.636981 ----------------------- -------------------------------------------------------- -Time: 1.022s Load: 0.017s, Pack+Encode: 0.415s, Decode+Unpack: 0.590s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 315, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 315, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6370 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample100-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample100-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample102-layer4-item1.zst (5/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample102-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 362, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 362, 128) -Output shape: (1, 362, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.0.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.1.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.1.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.2.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.2.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.3.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.3.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.4.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.4.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.4.output: torch.Size([1, 362, 4096]) -> torch.Size([1, 1, 362, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,012B, BPFP=0.1945 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,460B, BPFP=1.2832 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,948B, BPFP=0.6247 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,508B, BPFP=1.3922 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,940B, BPFP=0.8188 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,300B, BPFP=1.4093 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,100B, BPFP=0.8654 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,528B, BPFP=1.3710 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,160B, BPFP=0.6077 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,276B, BPFP=1.4303 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,400B, BPFP=0.8007 -⌛️ [2/4] FRONTEND: Frontend time: 0.410s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 362, 128]) - layer.0.v_cache: torch.Size([1, 8, 362, 128]) - layer.1.k_cache: torch.Size([1, 8, 362, 128]) - layer.1.v_cache: torch.Size([1, 8, 362, 128]) - layer.2.k_cache: torch.Size([1, 8, 362, 128]) - layer.2.v_cache: torch.Size([1, 8, 362, 128]) - layer.3.k_cache: torch.Size([1, 8, 362, 128]) - layer.3.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.k_cache: torch.Size([1, 8, 362, 128]) - layer.4.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.output: torch.Size([1, 362, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.684s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 362, 128]) - layer.0.v_cache: torch.Size([1, 8, 362, 128]) - layer.1.k_cache: torch.Size([1, 8, 362, 128]) - layer.1.v_cache: torch.Size([1, 8, 362, 128]) - layer.2.k_cache: torch.Size([1, 8, 362, 128]) - layer.2.v_cache: torch.Size([1, 8, 362, 128]) - layer.3.k_cache: torch.Size([1, 8, 362, 128]) - layer.3.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.k_cache: torch.Size([1, 8, 362, 128]) - layer.4.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.output: torch.Size([1, 362, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02398593 19.32142799 - layer.0.v_cache 0.00000027 0.00030568 - layer.1.k_cache 0.00294127 1.23796849 - layer.1.v_cache 0.00000079 0.00114442 - layer.2.k_cache 0.00116280 0.69572120 - layer.2.v_cache 0.00000114 0.00166387 - layer.3.k_cache 0.00131002 0.74949924 - layer.3.v_cache 0.00000215 0.00271136 - layer.4.k_cache 0.00366934 1.38667601 - layer.4.v_cache 0.00000328 0.00459020 - layer.4.output 0.00013411 0.08680771 - ------------------------------------------------------------------------------------- - TOTAL 0.00240096 1.69635281 - (elements=5,189,632) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5189632 -Total Bytes 611632 -BPFP 0.9429 bits/point -EBPFP 1.8857 equivalent bits/point -MSE 1.696353 ----------------------- -------------------------------------------------------- -Time: 1.114s Load: 0.019s, Pack+Encode: 0.410s, Decode+Unpack: 0.684s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 362, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6964 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample102-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample102-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample105-layer4-item1.zst (6/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample105-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 336, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 336, 128) -Output shape: (1, 336, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.output: torch.Size([1, 336, 4096]) -> torch.Size([1, 1, 336, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,084B, BPFP=0.2112 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,624B, BPFP=1.3398 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,844B, BPFP=0.6474 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,908B, BPFP=1.4395 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,540B, BPFP=0.8264 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,688B, BPFP=1.4808 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,976B, BPFP=0.9062 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,192B, BPFP=1.4461 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,364B, BPFP=0.6363 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,960B, BPFP=1.4872 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,836B, BPFP=0.8652 -⌛️ [2/4] FRONTEND: Frontend time: 0.409s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.output: torch.Size([1, 336, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.673s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.output: torch.Size([1, 336, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02466884 18.59171404 - layer.0.v_cache 0.00000026 0.00029967 - layer.1.k_cache 0.00293366 1.23805164 - layer.1.v_cache 0.00000078 0.00111569 - layer.2.k_cache 0.00117059 0.69215302 - layer.2.v_cache 0.00000113 0.00163777 - layer.3.k_cache 0.00132874 0.75790691 - layer.3.v_cache 0.00000215 0.00271052 - layer.4.k_cache 0.00360252 1.42946697 - layer.4.v_cache 0.00000314 0.00437679 - layer.4.output 0.00014135 0.08869732 - ------------------------------------------------------------------------------------- - TOTAL 0.00244837 1.64815874 - (elements=4,816,896) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4816896 -Total Bytes 597016 -BPFP 0.9915 bits/point -EBPFP 1.9831 equivalent bits/point -MSE 1.648159 ----------------------- -------------------------------------------------------- -Time: 1.098s Load: 0.017s, Pack+Encode: 0.409s, Decode+Unpack: 0.673s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 336, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6482 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample105-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample105-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample108-layer4-item1.zst (7/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample108-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 317, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 317, 128) -Output shape: (1, 317, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.0.v_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.1.k_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.1.v_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.2.k_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.2.v_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.3.k_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.3.v_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.4.k_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.4.v_cache: torch.Size([1, 8, 317, 128]) -> torch.Size([1, 1, 317, 1024]) - layer.4.output: torch.Size([1, 317, 4096]) -> torch.Size([1, 1, 317, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,548B, BPFP=0.2107 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,404B, BPFP=1.2422 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,600B, BPFP=0.6063 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 54,948B, BPFP=1.3542 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,608B, BPFP=0.8036 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,888B, BPFP=1.3774 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,200B, BPFP=0.8429 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 54,036B, BPFP=1.3317 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,164B, BPFP=0.5955 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,640B, BPFP=1.3959 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 131,184B, BPFP=0.8083 -⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 317, 128]) - layer.0.v_cache: torch.Size([1, 8, 317, 128]) - layer.1.k_cache: torch.Size([1, 8, 317, 128]) - layer.1.v_cache: torch.Size([1, 8, 317, 128]) - layer.2.k_cache: torch.Size([1, 8, 317, 128]) - layer.2.v_cache: torch.Size([1, 8, 317, 128]) - layer.3.k_cache: torch.Size([1, 8, 317, 128]) - layer.3.v_cache: torch.Size([1, 8, 317, 128]) - layer.4.k_cache: torch.Size([1, 8, 317, 128]) - layer.4.v_cache: torch.Size([1, 8, 317, 128]) - layer.4.output: torch.Size([1, 317, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.573s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 317, 128]) - layer.0.v_cache: torch.Size([1, 8, 317, 128]) - layer.1.k_cache: torch.Size([1, 8, 317, 128]) - layer.1.v_cache: torch.Size([1, 8, 317, 128]) - layer.2.k_cache: torch.Size([1, 8, 317, 128]) - layer.2.v_cache: torch.Size([1, 8, 317, 128]) - layer.3.k_cache: torch.Size([1, 8, 317, 128]) - layer.3.v_cache: torch.Size([1, 8, 317, 128]) - layer.4.k_cache: torch.Size([1, 8, 317, 128]) - layer.4.v_cache: torch.Size([1, 8, 317, 128]) - layer.4.output: torch.Size([1, 317, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02499472 17.78796579 - layer.0.v_cache 0.00000027 0.00030086 - layer.1.k_cache 0.00294942 1.13147801 - layer.1.v_cache 0.00000081 0.00114595 - layer.2.k_cache 0.00116222 0.68701427 - layer.2.v_cache 0.00000115 0.00170187 - layer.3.k_cache 0.00131165 0.75243298 - layer.3.v_cache 0.00000215 0.00279582 - layer.4.k_cache 0.00349495 1.37792661 - layer.4.v_cache 0.00000332 0.00464980 - layer.4.output 0.00014092 0.08898405 - ------------------------------------------------------------------------------------- - TOTAL 0.00246317 1.57881058 - (elements=4,544,512) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4544512 -Total Bytes 527220 -BPFP 0.9281 bits/point -EBPFP 1.8562 equivalent bits/point -MSE 1.578811 ----------------------- -------------------------------------------------------- -Time: 0.939s Load: 0.016s, Pack+Encode: 0.350s, Decode+Unpack: 0.573s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 317, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 317, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5788 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample108-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample108-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample11-layer4-item1.zst (8/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample11-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 353, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 353, 128) -Output shape: (1, 353, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.0.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.1.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.1.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.2.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.2.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.3.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.3.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.4.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.4.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.4.output: torch.Size([1, 353, 4096]) -> torch.Size([1, 1, 353, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,892B, BPFP=0.1968 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,048B, BPFP=1.3068 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,580B, BPFP=0.6104 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,564B, BPFP=1.4068 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,980B, BPFP=0.7963 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,916B, BPFP=1.4146 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,784B, BPFP=0.8584 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,912B, BPFP=1.3702 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,660B, BPFP=0.5900 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,588B, BPFP=1.4294 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 137,532B, BPFP=0.7610 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 353, 128]) - layer.0.v_cache: torch.Size([1, 8, 353, 128]) - layer.1.k_cache: torch.Size([1, 8, 353, 128]) - layer.1.v_cache: torch.Size([1, 8, 353, 128]) - layer.2.k_cache: torch.Size([1, 8, 353, 128]) - layer.2.v_cache: torch.Size([1, 8, 353, 128]) - layer.3.k_cache: torch.Size([1, 8, 353, 128]) - layer.3.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.k_cache: torch.Size([1, 8, 353, 128]) - layer.4.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.output: torch.Size([1, 353, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.671s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 353, 128]) - layer.0.v_cache: torch.Size([1, 8, 353, 128]) - layer.1.k_cache: torch.Size([1, 8, 353, 128]) - layer.1.v_cache: torch.Size([1, 8, 353, 128]) - layer.2.k_cache: torch.Size([1, 8, 353, 128]) - layer.2.v_cache: torch.Size([1, 8, 353, 128]) - layer.3.k_cache: torch.Size([1, 8, 353, 128]) - layer.3.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.k_cache: torch.Size([1, 8, 353, 128]) - layer.4.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.output: torch.Size([1, 353, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02498756 18.48560884 - layer.0.v_cache 0.00000027 0.00031149 - layer.1.k_cache 0.00288490 1.24075793 - layer.1.v_cache 0.00000079 0.00117829 - layer.2.k_cache 0.00118233 0.70395429 - layer.2.v_cache 0.00000115 0.00171298 - layer.3.k_cache 0.00132883 0.76427410 - layer.3.v_cache 0.00000217 0.00268054 - layer.4.k_cache 0.00368366 1.41445006 - layer.4.v_cache 0.00000317 0.00450779 - layer.4.output 0.00013605 0.08805406 - ------------------------------------------------------------------------------------- - TOTAL 0.00247279 1.64083233 - (elements=5,060,608) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5060608 -Total Bytes 588456 -BPFP 0.9303 bits/point -EBPFP 1.8605 equivalent bits/point -MSE 1.640832 ----------------------- -------------------------------------------------------- -Time: 1.092s Load: 0.018s, Pack+Encode: 0.402s, Decode+Unpack: 0.671s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 353, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6408 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample11-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample11-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample112-layer4-item1.zst (9/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample112-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 318, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.015s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 318, 128) -Output shape: (1, 318, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.0.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.1.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.1.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.2.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.2.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.3.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.3.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.4.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.4.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.4.output: torch.Size([1, 318, 4096]) -> torch.Size([1, 1, 318, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,596B, BPFP=0.2112 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,316B, BPFP=1.2361 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,460B, BPFP=0.6009 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,416B, BPFP=1.3123 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,628B, BPFP=0.7770 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 54,424B, BPFP=1.3371 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,124B, BPFP=0.8383 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,348B, BPFP=1.3106 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,660B, BPFP=0.5813 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,256B, BPFP=1.3575 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 123,816B, BPFP=0.7605 -⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 318, 128]) - layer.0.v_cache: torch.Size([1, 8, 318, 128]) - layer.1.k_cache: torch.Size([1, 8, 318, 128]) - layer.1.v_cache: torch.Size([1, 8, 318, 128]) - layer.2.k_cache: torch.Size([1, 8, 318, 128]) - layer.2.v_cache: torch.Size([1, 8, 318, 128]) - layer.3.k_cache: torch.Size([1, 8, 318, 128]) - layer.3.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.k_cache: torch.Size([1, 8, 318, 128]) - layer.4.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.output: torch.Size([1, 318, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.578s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 318, 128]) - layer.0.v_cache: torch.Size([1, 8, 318, 128]) - layer.1.k_cache: torch.Size([1, 8, 318, 128]) - layer.1.v_cache: torch.Size([1, 8, 318, 128]) - layer.2.k_cache: torch.Size([1, 8, 318, 128]) - layer.2.v_cache: torch.Size([1, 8, 318, 128]) - layer.3.k_cache: torch.Size([1, 8, 318, 128]) - layer.3.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.k_cache: torch.Size([1, 8, 318, 128]) - layer.4.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.output: torch.Size([1, 318, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02461332 18.86868766 - layer.0.v_cache 0.00000025 0.00028900 - layer.1.k_cache 0.00290030 1.22812725 - layer.1.v_cache 0.00000077 0.00104050 - layer.2.k_cache 0.00116467 0.68781674 - layer.2.v_cache 0.00000108 0.00153246 - layer.3.k_cache 0.00132360 0.75795386 - layer.3.v_cache 0.00000220 0.00258655 - layer.4.k_cache 0.00367546 1.42859146 - layer.4.v_cache 0.00000314 0.00435975 - layer.4.output 0.00019099 0.09532112 - ------------------------------------------------------------------------------------- - TOTAL 0.00246062 1.66873355 - (elements=4,558,848) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4558848 -Total Bytes 513044 -BPFP 0.9003 bits/point -EBPFP 1.8006 equivalent bits/point -MSE 1.668734 ----------------------- -------------------------------------------------------- -Time: 0.944s Load: 0.015s, Pack+Encode: 0.350s, Decode+Unpack: 0.578s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 318, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6687 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample112-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample112-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample114-layer4-item1.zst (10/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample114-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 330, 128) -Output shape: (1, 330, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.output: torch.Size([1, 330, 4096]) -> torch.Size([1, 1, 330, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,124B, BPFP=0.2160 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,496B, BPFP=1.3612 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,680B, BPFP=0.6790 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,332B, BPFP=1.4757 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,288B, BPFP=0.8591 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,520B, BPFP=1.5275 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,248B, BPFP=0.9055 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,516B, BPFP=1.4800 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,684B, BPFP=0.6554 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,680B, BPFP=1.5312 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,780B, BPFP=0.8806 -⌛️ [2/4] FRONTEND: Frontend time: 0.404s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.674s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02377391 20.09272461 - layer.0.v_cache 0.00000027 0.00030987 - layer.1.k_cache 0.00289258 1.19350133 - layer.1.v_cache 0.00000083 0.00113934 - layer.2.k_cache 0.00117497 0.69522418 - layer.2.v_cache 0.00000126 0.00171078 - layer.3.k_cache 0.00131231 0.75197620 - layer.3.v_cache 0.00000230 0.00280643 - layer.4.k_cache 0.00357793 1.38518261 - layer.4.v_cache 0.00000347 0.00472878 - layer.4.output 0.00014070 0.09263236 - ------------------------------------------------------------------------------------- - TOTAL 0.00237876 1.74998811 - (elements=4,730,880) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4730880 -Total Bytes 600348 -BPFP 1.0152 bits/point -EBPFP 2.0304 equivalent bits/point -MSE 1.749988 ----------------------- -------------------------------------------------------- -Time: 1.096s Load: 0.018s, Pack+Encode: 0.404s, Decode+Unpack: 0.674s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7500 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample114-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample114-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample119-layer4-item1.zst (11/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample119-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 357, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 357, 128) -Output shape: (1, 357, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.0.v_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.1.k_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.1.v_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.2.k_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.2.v_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.3.k_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.3.v_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.4.k_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.4.v_cache: torch.Size([1, 8, 357, 128]) -> torch.Size([1, 1, 357, 1024]) - layer.4.output: torch.Size([1, 357, 4096]) -> torch.Size([1, 1, 357, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,272B, BPFP=0.2029 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 60,684B, BPFP=1.3280 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,336B, BPFP=0.6420 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,132B, BPFP=1.4034 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,940B, BPFP=0.8084 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,616B, BPFP=1.4140 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,532B, BPFP=0.8870 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,308B, BPFP=1.3854 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,168B, BPFP=0.6164 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,840B, BPFP=1.4408 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 155,808B, BPFP=0.8524 -⌛️ [2/4] FRONTEND: Frontend time: 0.405s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 357, 128]) - layer.0.v_cache: torch.Size([1, 8, 357, 128]) - layer.1.k_cache: torch.Size([1, 8, 357, 128]) - layer.1.v_cache: torch.Size([1, 8, 357, 128]) - layer.2.k_cache: torch.Size([1, 8, 357, 128]) - layer.2.v_cache: torch.Size([1, 8, 357, 128]) - layer.3.k_cache: torch.Size([1, 8, 357, 128]) - layer.3.v_cache: torch.Size([1, 8, 357, 128]) - layer.4.k_cache: torch.Size([1, 8, 357, 128]) - layer.4.v_cache: torch.Size([1, 8, 357, 128]) - layer.4.output: torch.Size([1, 357, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.679s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 357, 128]) - layer.0.v_cache: torch.Size([1, 8, 357, 128]) - layer.1.k_cache: torch.Size([1, 8, 357, 128]) - layer.1.v_cache: torch.Size([1, 8, 357, 128]) - layer.2.k_cache: torch.Size([1, 8, 357, 128]) - layer.2.v_cache: torch.Size([1, 8, 357, 128]) - layer.3.k_cache: torch.Size([1, 8, 357, 128]) - layer.3.v_cache: torch.Size([1, 8, 357, 128]) - layer.4.k_cache: torch.Size([1, 8, 357, 128]) - layer.4.v_cache: torch.Size([1, 8, 357, 128]) - layer.4.output: torch.Size([1, 357, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02468803 19.15017179 - layer.0.v_cache 0.00000027 0.00031221 - layer.1.k_cache 0.00290374 1.23959513 - layer.1.v_cache 0.00000082 0.00114129 - layer.2.k_cache 0.00112985 0.68652087 - layer.2.v_cache 0.00000115 0.00160143 - layer.3.k_cache 0.00131122 0.76087443 - layer.3.v_cache 0.00000220 0.00274089 - layer.4.k_cache 0.00345488 1.40673435 - layer.4.v_cache 0.00000325 0.00454202 - layer.4.output 0.00016267 0.09711879 - ------------------------------------------------------------------------------------- - TOTAL 0.00243901 1.68876497 - (elements=5,117,952) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5117952 -Total Bytes 618636 -BPFP 0.9670 bits/point -EBPFP 1.9340 equivalent bits/point -MSE 1.688765 ----------------------- -------------------------------------------------------- -Time: 1.103s Load: 0.019s, Pack+Encode: 0.405s, Decode+Unpack: 0.679s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 357, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 357, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6888 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample119-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample119-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample12-layer4-item1.zst (12/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample12-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 358, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 358, 128) -Output shape: (1, 358, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.0.v_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.1.k_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.1.v_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.2.k_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.2.v_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.3.k_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.3.v_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.4.k_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.4.v_cache: torch.Size([1, 8, 358, 128]) -> torch.Size([1, 1, 358, 1024]) - layer.4.output: torch.Size([1, 358, 4096]) -> torch.Size([1, 1, 358, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,160B, BPFP=0.1999 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,540B, BPFP=1.2993 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,008B, BPFP=0.6330 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,768B, BPFP=1.3916 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,056B, BPFP=0.8087 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,516B, BPFP=1.4079 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,676B, BPFP=0.8658 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,084B, BPFP=1.3767 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,196B, BPFP=0.6153 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,496B, BPFP=1.4293 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,868B, BPFP=0.7904 -⌛️ [2/4] FRONTEND: Frontend time: 0.415s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 358, 128]) - layer.0.v_cache: torch.Size([1, 8, 358, 128]) - layer.1.k_cache: torch.Size([1, 8, 358, 128]) - layer.1.v_cache: torch.Size([1, 8, 358, 128]) - layer.2.k_cache: torch.Size([1, 8, 358, 128]) - layer.2.v_cache: torch.Size([1, 8, 358, 128]) - layer.3.k_cache: torch.Size([1, 8, 358, 128]) - layer.3.v_cache: torch.Size([1, 8, 358, 128]) - layer.4.k_cache: torch.Size([1, 8, 358, 128]) - layer.4.v_cache: torch.Size([1, 8, 358, 128]) - layer.4.output: torch.Size([1, 358, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 358, 128]) - layer.0.v_cache: torch.Size([1, 8, 358, 128]) - layer.1.k_cache: torch.Size([1, 8, 358, 128]) - layer.1.v_cache: torch.Size([1, 8, 358, 128]) - layer.2.k_cache: torch.Size([1, 8, 358, 128]) - layer.2.v_cache: torch.Size([1, 8, 358, 128]) - layer.3.k_cache: torch.Size([1, 8, 358, 128]) - layer.3.v_cache: torch.Size([1, 8, 358, 128]) - layer.4.k_cache: torch.Size([1, 8, 358, 128]) - layer.4.v_cache: torch.Size([1, 8, 358, 128]) - layer.4.output: torch.Size([1, 358, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02418174 19.18352146 - layer.0.v_cache 0.00000026 0.00030132 - layer.1.k_cache 0.00288075 1.20030519 - layer.1.v_cache 0.00000077 0.00111406 - layer.2.k_cache 0.00118336 0.68207699 - layer.2.v_cache 0.00000117 0.00169205 - layer.3.k_cache 0.00130988 0.75370175 - layer.3.v_cache 0.00000211 0.00272758 - layer.4.k_cache 0.00353008 1.39266115 - layer.4.v_cache 0.00000317 0.00463931 - layer.4.output 0.00013834 0.09073612 - ------------------------------------------------------------------------------------- - TOTAL 0.00240333 1.68469181 - (elements=5,132,288) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5132288 -Total Bytes 604368 -BPFP 0.9421 bits/point -EBPFP 1.8841 equivalent bits/point -MSE 1.684692 ----------------------- -------------------------------------------------------- -Time: 1.112s Load: 0.019s, Pack+Encode: 0.415s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 358, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 358, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6847 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample12-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample12-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample13-layer4-item1.zst (13/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample13-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 336, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 336, 128) -Output shape: (1, 336, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.output: torch.Size([1, 336, 4096]) -> torch.Size([1, 1, 336, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,124B, BPFP=0.2121 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,664B, BPFP=1.3408 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,216B, BPFP=0.6561 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,612B, BPFP=1.4558 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,136B, BPFP=0.8402 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,304B, BPFP=1.4952 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,284B, BPFP=0.9134 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,388B, BPFP=1.4506 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,972B, BPFP=0.6271 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,484B, BPFP=1.4993 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,956B, BPFP=0.8426 -⌛️ [2/4] FRONTEND: Frontend time: 0.403s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.output: torch.Size([1, 336, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.676s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.output: torch.Size([1, 336, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02407829 17.92127046 - layer.0.v_cache 0.00000026 0.00029579 - layer.1.k_cache 0.00285290 1.22240248 - layer.1.v_cache 0.00000077 0.00113437 - layer.2.k_cache 0.00116461 0.69726862 - layer.2.v_cache 0.00000115 0.00173371 - layer.3.k_cache 0.00131253 0.76522745 - layer.3.v_cache 0.00000220 0.00280192 - layer.4.k_cache 0.00369396 1.42355928 - layer.4.v_cache 0.00000328 0.00468456 - layer.4.output 0.00015589 0.09665569 - ------------------------------------------------------------------------------------- - TOTAL 0.00240954 1.60192867 - (elements=4,816,896) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4816896 -Total Bytes 596140 -BPFP 0.9901 bits/point -EBPFP 1.9802 equivalent bits/point -MSE 1.601929 ----------------------- -------------------------------------------------------- -Time: 1.095s Load: 0.016s, Pack+Encode: 0.403s, Decode+Unpack: 0.676s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 336, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6019 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample13-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample13-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample132-layer4-item1.zst (14/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample132-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 346, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 346, 128) -Output shape: (1, 346, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.0.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.1.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.1.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.2.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.2.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.3.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.3.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.4.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.4.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.4.output: torch.Size([1, 346, 4096]) -> torch.Size([1, 1, 346, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,188B, BPFP=0.2075 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,124B, BPFP=1.3350 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,664B, BPFP=0.6246 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,380B, BPFP=1.4311 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,328B, BPFP=0.8203 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,484B, BPFP=1.4560 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,256B, BPFP=0.8864 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,452B, BPFP=1.4327 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,792B, BPFP=0.6049 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,524B, BPFP=1.4795 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 160,692B, BPFP=0.9071 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 346, 128]) - layer.0.v_cache: torch.Size([1, 8, 346, 128]) - layer.1.k_cache: torch.Size([1, 8, 346, 128]) - layer.1.v_cache: torch.Size([1, 8, 346, 128]) - layer.2.k_cache: torch.Size([1, 8, 346, 128]) - layer.2.v_cache: torch.Size([1, 8, 346, 128]) - layer.3.k_cache: torch.Size([1, 8, 346, 128]) - layer.3.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.k_cache: torch.Size([1, 8, 346, 128]) - layer.4.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.output: torch.Size([1, 346, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.676s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 346, 128]) - layer.0.v_cache: torch.Size([1, 8, 346, 128]) - layer.1.k_cache: torch.Size([1, 8, 346, 128]) - layer.1.v_cache: torch.Size([1, 8, 346, 128]) - layer.2.k_cache: torch.Size([1, 8, 346, 128]) - layer.2.v_cache: torch.Size([1, 8, 346, 128]) - layer.3.k_cache: torch.Size([1, 8, 346, 128]) - layer.3.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.k_cache: torch.Size([1, 8, 346, 128]) - layer.4.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.output: torch.Size([1, 346, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02507911 19.60814583 - layer.0.v_cache 0.00000027 0.00030756 - layer.1.k_cache 0.00290669 1.29126241 - layer.1.v_cache 0.00000082 0.00113230 - layer.2.k_cache 0.00115107 0.68491531 - layer.2.v_cache 0.00000117 0.00166145 - layer.3.k_cache 0.00130922 0.75575512 - layer.3.v_cache 0.00000214 0.00274027 - layer.4.k_cache 0.00347357 1.43391621 - layer.4.v_cache 0.00000325 0.00453515 - layer.4.output 0.00015455 0.09851456 - ------------------------------------------------------------------------------------- - TOTAL 0.00246754 1.72703070 - (elements=4,960,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4960256 -Total Bytes 615884 -BPFP 0.9933 bits/point -EBPFP 1.9866 equivalent bits/point -MSE 1.727031 ----------------------- -------------------------------------------------------- -Time: 1.101s Load: 0.019s, Pack+Encode: 0.406s, Decode+Unpack: 0.676s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 346, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7270 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample132-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample132-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample135-layer4-item1.zst (15/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample135-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 343, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 343, 128) -Output shape: (1, 343, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.output: torch.Size([1, 343, 4096]) -> torch.Size([1, 1, 343, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,920B, BPFP=0.2032 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,912B, BPFP=1.3418 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,996B, BPFP=0.6377 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,736B, BPFP=1.4517 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,760B, BPFP=0.8373 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,516B, BPFP=1.4695 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,532B, BPFP=0.9004 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,320B, BPFP=1.4422 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,424B, BPFP=0.6246 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,440B, BPFP=1.4905 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 140,404B, BPFP=0.7995 -⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.output: torch.Size([1, 343, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.672s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.output: torch.Size([1, 343, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02309080 18.25296243 - layer.0.v_cache 0.00000026 0.00029167 - layer.1.k_cache 0.00292563 1.25082362 - layer.1.v_cache 0.00000076 0.00111069 - layer.2.k_cache 0.00117950 0.69168491 - layer.2.v_cache 0.00000116 0.00158829 - layer.3.k_cache 0.00131686 0.75702652 - layer.3.v_cache 0.00000208 0.00258256 - layer.4.k_cache 0.00357410 1.40659477 - layer.4.v_cache 0.00000334 0.00435355 - layer.4.output 0.00016206 0.09196086 - ------------------------------------------------------------------------------------- - TOTAL 0.00233876 1.62406160 - (elements=4,917,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4917248 -Total Bytes 596960 -BPFP 0.9712 bits/point -EBPFP 1.9424 equivalent bits/point -MSE 1.624062 ----------------------- -------------------------------------------------------- -Time: 1.090s Load: 0.017s, Pack+Encode: 0.400s, Decode+Unpack: 0.672s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 343, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6241 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample135-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample135-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample14-layer4-item1.zst (16/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample14-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 359, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 359, 128) -Output shape: (1, 359, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.0.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.1.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.1.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.2.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.2.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.3.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.3.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.4.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.4.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.4.output: torch.Size([1, 359, 4096]) -> torch.Size([1, 1, 359, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,096B, BPFP=0.1979 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,556B, BPFP=1.2960 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,644B, BPFP=0.6451 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 65,140B, BPFP=1.4176 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 38,028B, BPFP=0.8276 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,392B, BPFP=1.4231 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,184B, BPFP=0.8745 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,944B, BPFP=1.3915 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,604B, BPFP=0.6225 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,064B, BPFP=1.4377 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 151,696B, BPFP=0.8253 -⌛️ [2/4] FRONTEND: Frontend time: 0.399s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 359, 128]) - layer.0.v_cache: torch.Size([1, 8, 359, 128]) - layer.1.k_cache: torch.Size([1, 8, 359, 128]) - layer.1.v_cache: torch.Size([1, 8, 359, 128]) - layer.2.k_cache: torch.Size([1, 8, 359, 128]) - layer.2.v_cache: torch.Size([1, 8, 359, 128]) - layer.3.k_cache: torch.Size([1, 8, 359, 128]) - layer.3.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.k_cache: torch.Size([1, 8, 359, 128]) - layer.4.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.output: torch.Size([1, 359, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.681s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 359, 128]) - layer.0.v_cache: torch.Size([1, 8, 359, 128]) - layer.1.k_cache: torch.Size([1, 8, 359, 128]) - layer.1.v_cache: torch.Size([1, 8, 359, 128]) - layer.2.k_cache: torch.Size([1, 8, 359, 128]) - layer.2.v_cache: torch.Size([1, 8, 359, 128]) - layer.3.k_cache: torch.Size([1, 8, 359, 128]) - layer.3.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.k_cache: torch.Size([1, 8, 359, 128]) - layer.4.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.output: torch.Size([1, 359, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02394788 19.08200269 - layer.0.v_cache 0.00000027 0.00030049 - layer.1.k_cache 0.00287336 1.18382420 - layer.1.v_cache 0.00000081 0.00120204 - layer.2.k_cache 0.00118844 0.69791848 - layer.2.v_cache 0.00000120 0.00174927 - layer.3.k_cache 0.00133454 0.76404701 - layer.3.v_cache 0.00000220 0.00277719 - layer.4.k_cache 0.00356152 1.37784647 - layer.4.v_cache 0.00000363 0.00476502 - layer.4.output 0.00015158 0.09147069 - ------------------------------------------------------------------------------------- - TOTAL 0.00239430 1.67730826 - (elements=5,146,624) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5146624 -Total Bytes 617348 -BPFP 0.9596 bits/point -EBPFP 1.9192 equivalent bits/point -MSE 1.677308 ----------------------- -------------------------------------------------------- -Time: 1.098s Load: 0.018s, Pack+Encode: 0.399s, Decode+Unpack: 0.681s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 359, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6773 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample14-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample14-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample15-layer4-item1.zst (17/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample15-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 359, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 359, 128) -Output shape: (1, 359, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.0.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.1.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.1.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.2.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.2.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.3.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.3.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.4.k_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.4.v_cache: torch.Size([1, 8, 359, 128]) -> torch.Size([1, 1, 359, 1024]) - layer.4.output: torch.Size([1, 359, 4096]) -> torch.Size([1, 1, 359, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,224B, BPFP=0.2007 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,456B, BPFP=1.2939 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,180B, BPFP=0.6350 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,336B, BPFP=1.4001 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 38,108B, BPFP=0.8293 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,252B, BPFP=1.4200 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,456B, BPFP=0.8804 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,716B, BPFP=1.3866 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,360B, BPFP=0.6172 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,036B, BPFP=1.4371 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,532B, BPFP=0.7918 -⌛️ [2/4] FRONTEND: Frontend time: 0.410s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 359, 128]) - layer.0.v_cache: torch.Size([1, 8, 359, 128]) - layer.1.k_cache: torch.Size([1, 8, 359, 128]) - layer.1.v_cache: torch.Size([1, 8, 359, 128]) - layer.2.k_cache: torch.Size([1, 8, 359, 128]) - layer.2.v_cache: torch.Size([1, 8, 359, 128]) - layer.3.k_cache: torch.Size([1, 8, 359, 128]) - layer.3.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.k_cache: torch.Size([1, 8, 359, 128]) - layer.4.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.output: torch.Size([1, 359, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.679s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 359, 128]) - layer.0.v_cache: torch.Size([1, 8, 359, 128]) - layer.1.k_cache: torch.Size([1, 8, 359, 128]) - layer.1.v_cache: torch.Size([1, 8, 359, 128]) - layer.2.k_cache: torch.Size([1, 8, 359, 128]) - layer.2.v_cache: torch.Size([1, 8, 359, 128]) - layer.3.k_cache: torch.Size([1, 8, 359, 128]) - layer.3.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.k_cache: torch.Size([1, 8, 359, 128]) - layer.4.v_cache: torch.Size([1, 8, 359, 128]) - layer.4.output: torch.Size([1, 359, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02338462 19.14172533 - layer.0.v_cache 0.00000027 0.00030711 - layer.1.k_cache 0.00293092 1.22496215 - layer.1.v_cache 0.00000080 0.00116177 - layer.2.k_cache 0.00116855 0.68950502 - layer.2.v_cache 0.00000117 0.00171826 - layer.3.k_cache 0.00131991 0.76403163 - layer.3.v_cache 0.00000231 0.00290173 - layer.4.k_cache 0.00358756 1.37358735 - layer.4.v_cache 0.00000356 0.00480655 - layer.4.output 0.00016308 0.08993328 - ------------------------------------------------------------------------------------- - TOTAL 0.00236086 1.68317429 - (elements=5,146,624) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5146624 -Total Bytes 609656 -BPFP 0.9477 bits/point -EBPFP 1.8953 equivalent bits/point -MSE 1.683174 ----------------------- -------------------------------------------------------- -Time: 1.107s Load: 0.018s, Pack+Encode: 0.410s, Decode+Unpack: 0.679s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 359, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 359, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6832 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample15-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample15-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample16-layer4-item1.zst (18/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample16-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 347, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 347, 128) -Output shape: (1, 347, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.0.v_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.1.k_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.1.v_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.2.k_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.2.v_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.3.k_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.3.v_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.4.k_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.4.v_cache: torch.Size([1, 8, 347, 128]) -> torch.Size([1, 1, 347, 1024]) - layer.4.output: torch.Size([1, 347, 4096]) -> torch.Size([1, 1, 347, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,872B, BPFP=0.1997 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,288B, BPFP=1.3348 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,228B, BPFP=0.6355 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,120B, BPFP=1.4436 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,012B, BPFP=0.8333 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,100B, BPFP=1.4657 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,640B, BPFP=0.8925 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,236B, BPFP=1.4237 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,296B, BPFP=0.6146 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,944B, BPFP=1.4847 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 142,592B, BPFP=0.8026 -⌛️ [2/4] FRONTEND: Frontend time: 0.403s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 347, 128]) - layer.0.v_cache: torch.Size([1, 8, 347, 128]) - layer.1.k_cache: torch.Size([1, 8, 347, 128]) - layer.1.v_cache: torch.Size([1, 8, 347, 128]) - layer.2.k_cache: torch.Size([1, 8, 347, 128]) - layer.2.v_cache: torch.Size([1, 8, 347, 128]) - layer.3.k_cache: torch.Size([1, 8, 347, 128]) - layer.3.v_cache: torch.Size([1, 8, 347, 128]) - layer.4.k_cache: torch.Size([1, 8, 347, 128]) - layer.4.v_cache: torch.Size([1, 8, 347, 128]) - layer.4.output: torch.Size([1, 347, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.681s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 347, 128]) - layer.0.v_cache: torch.Size([1, 8, 347, 128]) - layer.1.k_cache: torch.Size([1, 8, 347, 128]) - layer.1.v_cache: torch.Size([1, 8, 347, 128]) - layer.2.k_cache: torch.Size([1, 8, 347, 128]) - layer.2.v_cache: torch.Size([1, 8, 347, 128]) - layer.3.k_cache: torch.Size([1, 8, 347, 128]) - layer.3.v_cache: torch.Size([1, 8, 347, 128]) - layer.4.k_cache: torch.Size([1, 8, 347, 128]) - layer.4.v_cache: torch.Size([1, 8, 347, 128]) - layer.4.output: torch.Size([1, 347, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02341881 18.28782139 - layer.0.v_cache 0.00000026 0.00030829 - layer.1.k_cache 0.00293388 1.31008498 - layer.1.v_cache 0.00000079 0.00116388 - layer.2.k_cache 0.00115427 0.69288446 - layer.2.v_cache 0.00000115 0.00165370 - layer.3.k_cache 0.00129772 0.75333495 - layer.3.v_cache 0.00000217 0.00273420 - layer.4.k_cache 0.00353122 1.41038861 - layer.4.v_cache 0.00000331 0.00467884 - layer.4.output 0.00014441 0.09038991 - ------------------------------------------------------------------------------------- - TOTAL 0.00235152 1.63047235 - (elements=4,974,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4974592 -Total Bytes 601328 -BPFP 0.9670 bits/point -EBPFP 1.9341 equivalent bits/point -MSE 1.630472 ----------------------- -------------------------------------------------------- -Time: 1.101s Load: 0.017s, Pack+Encode: 0.403s, Decode+Unpack: 0.681s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 347, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 347, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6305 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample16-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample16-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample165-layer4-item1.zst (19/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample165-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 337, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 337, 128) -Output shape: (1, 337, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.0.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.1.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.1.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.2.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.2.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.3.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.3.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.4.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.4.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.4.output: torch.Size([1, 337, 4096]) -> torch.Size([1, 1, 337, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,876B, BPFP=0.2058 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,108B, BPFP=1.3239 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,380B, BPFP=0.6579 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 60,832B, BPFP=1.4102 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,392B, BPFP=0.8205 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,096B, BPFP=1.4627 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,168B, BPFP=0.9080 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,468B, BPFP=1.4250 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,192B, BPFP=0.6304 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,732B, BPFP=1.4775 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 137,252B, BPFP=0.7955 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 337, 128]) - layer.0.v_cache: torch.Size([1, 8, 337, 128]) - layer.1.k_cache: torch.Size([1, 8, 337, 128]) - layer.1.v_cache: torch.Size([1, 8, 337, 128]) - layer.2.k_cache: torch.Size([1, 8, 337, 128]) - layer.2.v_cache: torch.Size([1, 8, 337, 128]) - layer.3.k_cache: torch.Size([1, 8, 337, 128]) - layer.3.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.k_cache: torch.Size([1, 8, 337, 128]) - layer.4.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.output: torch.Size([1, 337, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.672s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 337, 128]) - layer.0.v_cache: torch.Size([1, 8, 337, 128]) - layer.1.k_cache: torch.Size([1, 8, 337, 128]) - layer.1.v_cache: torch.Size([1, 8, 337, 128]) - layer.2.k_cache: torch.Size([1, 8, 337, 128]) - layer.2.v_cache: torch.Size([1, 8, 337, 128]) - layer.3.k_cache: torch.Size([1, 8, 337, 128]) - layer.3.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.k_cache: torch.Size([1, 8, 337, 128]) - layer.4.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.output: torch.Size([1, 337, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02403486 18.78653706 - layer.0.v_cache 0.00000026 0.00030170 - layer.1.k_cache 0.00293885 1.24366448 - layer.1.v_cache 0.00000076 0.00103018 - layer.2.k_cache 0.00113696 0.69533894 - layer.2.v_cache 0.00000113 0.00150597 - layer.3.k_cache 0.00132482 0.78578897 - layer.3.v_cache 0.00000208 0.00256993 - layer.4.k_cache 0.00376518 1.48203647 - layer.4.v_cache 0.00000304 0.00431145 - layer.4.output 0.00014776 0.09604101 - ------------------------------------------------------------------------------------- - TOTAL 0.00241421 1.67051780 - (elements=4,831,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4831232 -Total Bytes 582496 -BPFP 0.9646 bits/point -EBPFP 1.9291 equivalent bits/point -MSE 1.670518 ----------------------- -------------------------------------------------------- -Time: 1.091s Load: 0.016s, Pack+Encode: 0.402s, Decode+Unpack: 0.672s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 337, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6705 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample165-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample165-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample17-layer4-item1.zst (20/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample17-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 351, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 351, 128) -Output shape: (1, 351, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.output: torch.Size([1, 351, 4096]) -> torch.Size([1, 1, 351, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,920B, BPFP=0.1985 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,500B, BPFP=1.3021 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,676B, BPFP=0.6160 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,308B, BPFP=1.3868 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,964B, BPFP=0.8005 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,224B, BPFP=1.4072 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,660B, BPFP=0.8605 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,160B, BPFP=1.3613 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,448B, BPFP=0.5887 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,804B, BPFP=1.4201 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,176B, BPFP=0.7856 -⌛️ [2/4] FRONTEND: Frontend time: 0.401s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.output: torch.Size([1, 351, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.677s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.output: torch.Size([1, 351, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02526469 18.87308444 - layer.0.v_cache 0.00000026 0.00030399 - layer.1.k_cache 0.00294286 1.19511009 - layer.1.v_cache 0.00000076 0.00109148 - layer.2.k_cache 0.00116394 0.69665375 - layer.2.v_cache 0.00000112 0.00158523 - layer.3.k_cache 0.00133268 0.75683837 - layer.3.v_cache 0.00000207 0.00257476 - layer.4.k_cache 0.00360177 1.38649895 - layer.4.v_cache 0.00000315 0.00437976 - layer.4.output 0.00019540 0.09245863 - ------------------------------------------------------------------------------------- - TOTAL 0.00250678 1.66342538 - (elements=5,031,936) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5031936 -Total Bytes 587840 -BPFP 0.9346 bits/point -EBPFP 1.8691 equivalent bits/point -MSE 1.663425 ----------------------- -------------------------------------------------------- -Time: 1.096s Load: 0.018s, Pack+Encode: 0.401s, Decode+Unpack: 0.677s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 351, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6634 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample17-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample17-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample18-layer4-item1.zst (21/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample18-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 346, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 346, 128) -Output shape: (1, 346, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.0.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.1.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.1.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.2.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.2.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.3.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.3.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.4.k_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.4.v_cache: torch.Size([1, 8, 346, 128]) -> torch.Size([1, 1, 346, 1024]) - layer.4.output: torch.Size([1, 346, 4096]) -> torch.Size([1, 1, 346, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,900B, BPFP=0.2010 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,032B, BPFP=1.3329 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,040B, BPFP=0.6331 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,980B, BPFP=1.4446 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,964B, BPFP=0.8346 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,452B, BPFP=1.4779 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,208B, BPFP=0.8853 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,600B, BPFP=1.4361 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,844B, BPFP=0.6061 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,844B, BPFP=1.4867 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,192B, BPFP=0.8196 -⌛️ [2/4] FRONTEND: Frontend time: 0.401s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 346, 128]) - layer.0.v_cache: torch.Size([1, 8, 346, 128]) - layer.1.k_cache: torch.Size([1, 8, 346, 128]) - layer.1.v_cache: torch.Size([1, 8, 346, 128]) - layer.2.k_cache: torch.Size([1, 8, 346, 128]) - layer.2.v_cache: torch.Size([1, 8, 346, 128]) - layer.3.k_cache: torch.Size([1, 8, 346, 128]) - layer.3.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.k_cache: torch.Size([1, 8, 346, 128]) - layer.4.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.output: torch.Size([1, 346, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.675s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 346, 128]) - layer.0.v_cache: torch.Size([1, 8, 346, 128]) - layer.1.k_cache: torch.Size([1, 8, 346, 128]) - layer.1.v_cache: torch.Size([1, 8, 346, 128]) - layer.2.k_cache: torch.Size([1, 8, 346, 128]) - layer.2.v_cache: torch.Size([1, 8, 346, 128]) - layer.3.k_cache: torch.Size([1, 8, 346, 128]) - layer.3.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.k_cache: torch.Size([1, 8, 346, 128]) - layer.4.v_cache: torch.Size([1, 8, 346, 128]) - layer.4.output: torch.Size([1, 346, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02293222 18.90770920 - layer.0.v_cache 0.00000026 0.00029096 - layer.1.k_cache 0.00295457 1.23352703 - layer.1.v_cache 0.00000081 0.00115156 - layer.2.k_cache 0.00115565 0.68110683 - layer.2.v_cache 0.00000119 0.00170415 - layer.3.k_cache 0.00131728 0.75245534 - layer.3.v_cache 0.00000219 0.00276652 - layer.4.k_cache 0.00358690 1.45136874 - layer.4.v_cache 0.00000349 0.00463160 - layer.4.output 0.00015659 0.09333353 - ------------------------------------------------------------------------------------- - TOTAL 0.00232721 1.67214615 - (elements=4,960,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4960256 -Total Bytes 603056 -BPFP 0.9726 bits/point -EBPFP 1.9452 equivalent bits/point -MSE 1.672146 ----------------------- -------------------------------------------------------- -Time: 1.092s Load: 0.016s, Pack+Encode: 0.401s, Decode+Unpack: 0.675s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 346, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 346, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6721 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample18-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample18-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample19-layer4-item1.zst (22/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample19-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 340, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 340, 128) -Output shape: (1, 340, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.0.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.1.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.1.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.2.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.2.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.3.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.3.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.4.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.4.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.4.output: torch.Size([1, 340, 4096]) -> torch.Size([1, 1, 340, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,468B, BPFP=0.2176 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,092B, BPFP=1.3578 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,308B, BPFP=0.6505 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,708B, BPFP=1.4639 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,656B, BPFP=0.8423 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,016B, BPFP=1.4939 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,100B, BPFP=0.8984 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,244B, BPFP=1.4532 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,356B, BPFP=0.6286 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,744B, BPFP=1.5107 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,048B, BPFP=0.8102 -⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 340, 128]) - layer.0.v_cache: torch.Size([1, 8, 340, 128]) - layer.1.k_cache: torch.Size([1, 8, 340, 128]) - layer.1.v_cache: torch.Size([1, 8, 340, 128]) - layer.2.k_cache: torch.Size([1, 8, 340, 128]) - layer.2.v_cache: torch.Size([1, 8, 340, 128]) - layer.3.k_cache: torch.Size([1, 8, 340, 128]) - layer.3.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.k_cache: torch.Size([1, 8, 340, 128]) - layer.4.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.output: torch.Size([1, 340, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.675s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 340, 128]) - layer.0.v_cache: torch.Size([1, 8, 340, 128]) - layer.1.k_cache: torch.Size([1, 8, 340, 128]) - layer.1.v_cache: torch.Size([1, 8, 340, 128]) - layer.2.k_cache: torch.Size([1, 8, 340, 128]) - layer.2.v_cache: torch.Size([1, 8, 340, 128]) - layer.3.k_cache: torch.Size([1, 8, 340, 128]) - layer.3.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.k_cache: torch.Size([1, 8, 340, 128]) - layer.4.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.output: torch.Size([1, 340, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02361420 19.25742906 - layer.0.v_cache 0.00000026 0.00031401 - layer.1.k_cache 0.00281368 1.18773193 - layer.1.v_cache 0.00000080 0.00118497 - layer.2.k_cache 0.00120688 0.71817555 - layer.2.v_cache 0.00000119 0.00176579 - layer.3.k_cache 0.00132000 0.76424875 - layer.3.v_cache 0.00000216 0.00280148 - layer.4.k_cache 0.00354241 1.38155482 - layer.4.v_cache 0.00000378 0.00483161 - layer.4.output 0.00015030 0.08784973 - ------------------------------------------------------------------------------------- - TOTAL 0.00236476 1.69081692 - (elements=4,874,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4874240 -Total Bytes 598740 -BPFP 0.9827 bits/point -EBPFP 1.9654 equivalent bits/point -MSE 1.690817 ----------------------- -------------------------------------------------------- -Time: 1.092s Load: 0.017s, Pack+Encode: 0.400s, Decode+Unpack: 0.675s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 340, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6908 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample19-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample19-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample2-layer4-item1.zst (23/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample2-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 344, 128) -Output shape: (1, 344, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.output: torch.Size([1, 344, 4096]) -> torch.Size([1, 1, 344, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,136B, BPFP=0.2075 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,148B, BPFP=1.3433 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,020B, BPFP=0.6364 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,596B, BPFP=1.4443 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,140B, BPFP=0.8208 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,268B, BPFP=1.4823 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,088B, BPFP=0.8877 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,192B, BPFP=1.4351 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,116B, BPFP=0.6158 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,168B, BPFP=1.5027 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,004B, BPFP=0.8233 -⌛️ [2/4] FRONTEND: Frontend time: 0.403s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02321494 19.98197334 - layer.0.v_cache 0.00000027 0.00028854 - layer.1.k_cache 0.00289516 1.19081125 - layer.1.v_cache 0.00000077 0.00110246 - layer.2.k_cache 0.00116570 0.68249055 - layer.2.v_cache 0.00000112 0.00161463 - layer.3.k_cache 0.00132982 0.73867620 - layer.3.v_cache 0.00000218 0.00268781 - layer.4.k_cache 0.00351457 1.39446977 - layer.4.v_cache 0.00000367 0.00472813 - layer.4.output 0.00015182 0.09318076 - ------------------------------------------------------------------------------------- - TOTAL 0.00233825 1.74082612 - (elements=4,931,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4931584 -Total Bytes 601876 -BPFP 0.9764 bits/point -EBPFP 1.9527 equivalent bits/point -MSE 1.740826 ----------------------- -------------------------------------------------------- -Time: 1.098s Load: 0.017s, Pack+Encode: 0.403s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7408 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample2-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample2-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample20-layer4-item1.zst (24/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample20-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 337, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 337, 128) -Output shape: (1, 337, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.0.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.1.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.1.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.2.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.2.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.3.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.3.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.4.k_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.4.v_cache: torch.Size([1, 8, 337, 128]) -> torch.Size([1, 1, 337, 1024]) - layer.4.output: torch.Size([1, 337, 4096]) -> torch.Size([1, 1, 337, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,932B, BPFP=0.2071 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,172B, BPFP=1.3486 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,128B, BPFP=0.6521 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,808B, BPFP=1.4560 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,732B, BPFP=0.8284 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,240B, BPFP=1.4892 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,748B, BPFP=0.8983 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,252B, BPFP=1.4432 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,764B, BPFP=0.6205 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,716B, BPFP=1.5003 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 138,752B, BPFP=0.8042 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 337, 128]) - layer.0.v_cache: torch.Size([1, 8, 337, 128]) - layer.1.k_cache: torch.Size([1, 8, 337, 128]) - layer.1.v_cache: torch.Size([1, 8, 337, 128]) - layer.2.k_cache: torch.Size([1, 8, 337, 128]) - layer.2.v_cache: torch.Size([1, 8, 337, 128]) - layer.3.k_cache: torch.Size([1, 8, 337, 128]) - layer.3.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.k_cache: torch.Size([1, 8, 337, 128]) - layer.4.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.output: torch.Size([1, 337, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.693s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 337, 128]) - layer.0.v_cache: torch.Size([1, 8, 337, 128]) - layer.1.k_cache: torch.Size([1, 8, 337, 128]) - layer.1.v_cache: torch.Size([1, 8, 337, 128]) - layer.2.k_cache: torch.Size([1, 8, 337, 128]) - layer.2.v_cache: torch.Size([1, 8, 337, 128]) - layer.3.k_cache: torch.Size([1, 8, 337, 128]) - layer.3.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.k_cache: torch.Size([1, 8, 337, 128]) - layer.4.v_cache: torch.Size([1, 8, 337, 128]) - layer.4.output: torch.Size([1, 337, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02435875 19.11818724 - layer.0.v_cache 0.00000026 0.00030883 - layer.1.k_cache 0.00286757 1.27705804 - layer.1.v_cache 0.00000078 0.00116243 - layer.2.k_cache 0.00117745 0.70361038 - layer.2.v_cache 0.00000112 0.00165644 - layer.3.k_cache 0.00130673 0.76744763 - layer.3.v_cache 0.00000209 0.00270282 - layer.4.k_cache 0.00356837 1.44599566 - layer.4.v_cache 0.00000322 0.00463115 - layer.4.output 0.00014638 0.09332240 - ------------------------------------------------------------------------------------- - TOTAL 0.00241942 1.69257502 - (elements=4,831,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4831232 -Total Bytes 589244 -BPFP 0.9757 bits/point -EBPFP 1.9514 equivalent bits/point -MSE 1.692575 ----------------------- -------------------------------------------------------- -Time: 1.117s Load: 0.017s, Pack+Encode: 0.407s, Decode+Unpack: 0.693s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 337, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 337, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6926 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample20-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample20-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample21-layer4-item1.zst (25/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample21-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 395, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.020s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 395, 128) -Output shape: (1, 395, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.0.v_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.1.k_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.1.v_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.2.k_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.2.v_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.3.k_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.3.v_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.4.k_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.4.v_cache: torch.Size([1, 8, 395, 128]) -> torch.Size([1, 1, 395, 1024]) - layer.4.output: torch.Size([1, 395, 4096]) -> torch.Size([1, 1, 395, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 10,268B, BPFP=0.2031 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,796B, BPFP=1.3211 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 32,108B, BPFP=0.6350 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,040B, BPFP=1.4248 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 40,348B, BPFP=0.7980 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,596B, BPFP=1.4556 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 44,484B, BPFP=0.8798 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,056B, BPFP=1.4252 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 31,620B, BPFP=0.6254 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,496B, BPFP=1.4734 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 168,616B, BPFP=0.8337 -⌛️ [2/4] FRONTEND: Frontend time: 0.615s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 395, 128]) - layer.0.v_cache: torch.Size([1, 8, 395, 128]) - layer.1.k_cache: torch.Size([1, 8, 395, 128]) - layer.1.v_cache: torch.Size([1, 8, 395, 128]) - layer.2.k_cache: torch.Size([1, 8, 395, 128]) - layer.2.v_cache: torch.Size([1, 8, 395, 128]) - layer.3.k_cache: torch.Size([1, 8, 395, 128]) - layer.3.v_cache: torch.Size([1, 8, 395, 128]) - layer.4.k_cache: torch.Size([1, 8, 395, 128]) - layer.4.v_cache: torch.Size([1, 8, 395, 128]) - layer.4.output: torch.Size([1, 395, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.782s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 395, 128]) - layer.0.v_cache: torch.Size([1, 8, 395, 128]) - layer.1.k_cache: torch.Size([1, 8, 395, 128]) - layer.1.v_cache: torch.Size([1, 8, 395, 128]) - layer.2.k_cache: torch.Size([1, 8, 395, 128]) - layer.2.v_cache: torch.Size([1, 8, 395, 128]) - layer.3.k_cache: torch.Size([1, 8, 395, 128]) - layer.3.v_cache: torch.Size([1, 8, 395, 128]) - layer.4.k_cache: torch.Size([1, 8, 395, 128]) - layer.4.v_cache: torch.Size([1, 8, 395, 128]) - layer.4.output: torch.Size([1, 395, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02333523 19.88912678 - layer.0.v_cache 0.00000027 0.00031600 - layer.1.k_cache 0.00291865 1.22281355 - layer.1.v_cache 0.00000080 0.00110832 - layer.2.k_cache 0.00114101 0.68776716 - layer.2.v_cache 0.00000113 0.00157157 - layer.3.k_cache 0.00133612 0.76507592 - layer.3.v_cache 0.00000221 0.00260506 - layer.4.k_cache 0.00365959 1.45331568 - layer.4.v_cache 0.00000309 0.00440270 - layer.4.output 0.00017803 0.09513709 - ------------------------------------------------------------------------------------- - TOTAL 0.00236502 1.74347508 - (elements=5,662,720) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5662720 -Total Bytes 686428 -BPFP 0.9698 bits/point -EBPFP 1.9395 equivalent bits/point -MSE 1.743475 ----------------------- -------------------------------------------------------- -Time: 1.416s Load: 0.020s, Pack+Encode: 0.615s, Decode+Unpack: 0.782s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 395, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 395, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7435 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample21-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample21-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample22-layer4-item1.zst (26/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample22-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 348, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 348, 128) -Output shape: (1, 348, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.0.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.1.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.1.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.2.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.2.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.3.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.3.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.4.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.4.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.4.output: torch.Size([1, 348, 4096]) -> torch.Size([1, 1, 348, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,012B, BPFP=0.2023 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,460B, BPFP=1.3349 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,140B, BPFP=0.6317 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,572B, BPFP=1.4272 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,456B, BPFP=0.8184 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,992B, BPFP=1.4591 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,572B, BPFP=0.8884 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,248B, BPFP=1.3974 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,752B, BPFP=0.6006 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,140B, BPFP=1.4624 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 136,464B, BPFP=0.7659 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 348, 128]) - layer.0.v_cache: torch.Size([1, 8, 348, 128]) - layer.1.k_cache: torch.Size([1, 8, 348, 128]) - layer.1.v_cache: torch.Size([1, 8, 348, 128]) - layer.2.k_cache: torch.Size([1, 8, 348, 128]) - layer.2.v_cache: torch.Size([1, 8, 348, 128]) - layer.3.k_cache: torch.Size([1, 8, 348, 128]) - layer.3.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.k_cache: torch.Size([1, 8, 348, 128]) - layer.4.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.output: torch.Size([1, 348, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.694s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 348, 128]) - layer.0.v_cache: torch.Size([1, 8, 348, 128]) - layer.1.k_cache: torch.Size([1, 8, 348, 128]) - layer.1.v_cache: torch.Size([1, 8, 348, 128]) - layer.2.k_cache: torch.Size([1, 8, 348, 128]) - layer.2.v_cache: torch.Size([1, 8, 348, 128]) - layer.3.k_cache: torch.Size([1, 8, 348, 128]) - layer.3.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.k_cache: torch.Size([1, 8, 348, 128]) - layer.4.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.output: torch.Size([1, 348, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02579716 19.43994000 - layer.0.v_cache 0.00000026 0.00031229 - layer.1.k_cache 0.00289751 1.24568132 - layer.1.v_cache 0.00000077 0.00115873 - layer.2.k_cache 0.00114660 0.69896650 - layer.2.v_cache 0.00000117 0.00169482 - layer.3.k_cache 0.00133042 0.76158502 - layer.3.v_cache 0.00000210 0.00271289 - layer.4.k_cache 0.00358099 1.46500458 - layer.4.v_cache 0.00000335 0.00469981 - layer.4.output 0.00013932 0.09088467 - ------------------------------------------------------------------------------------- - TOTAL 0.00252269 1.71323533 - (elements=4,988,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4988928 -Total Bytes 591808 -BPFP 0.9490 bits/point -EBPFP 1.8980 equivalent bits/point -MSE 1.713235 ----------------------- -------------------------------------------------------- -Time: 1.118s Load: 0.017s, Pack+Encode: 0.407s, Decode+Unpack: 0.694s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 348, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7132 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample22-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample22-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample23-layer4-item1.zst (27/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample23-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 316, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 316, 128) -Output shape: (1, 316, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.0.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.1.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.1.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.2.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.2.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.3.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.3.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.4.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.4.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.4.output: torch.Size([1, 316, 4096]) -> torch.Size([1, 1, 316, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,388B, BPFP=0.2074 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,660B, BPFP=1.2525 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,076B, BPFP=0.5952 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 55,032B, BPFP=1.3606 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,928B, BPFP=0.7894 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,516B, BPFP=1.3725 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,136B, BPFP=0.8439 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,948B, BPFP=1.3338 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,580B, BPFP=0.5830 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,440B, BPFP=1.3954 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 124,404B, BPFP=0.7689 -⌛️ [2/4] FRONTEND: Frontend time: 0.357s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 316, 128]) - layer.0.v_cache: torch.Size([1, 8, 316, 128]) - layer.1.k_cache: torch.Size([1, 8, 316, 128]) - layer.1.v_cache: torch.Size([1, 8, 316, 128]) - layer.2.k_cache: torch.Size([1, 8, 316, 128]) - layer.2.v_cache: torch.Size([1, 8, 316, 128]) - layer.3.k_cache: torch.Size([1, 8, 316, 128]) - layer.3.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.k_cache: torch.Size([1, 8, 316, 128]) - layer.4.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.output: torch.Size([1, 316, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.591s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 316, 128]) - layer.0.v_cache: torch.Size([1, 8, 316, 128]) - layer.1.k_cache: torch.Size([1, 8, 316, 128]) - layer.1.v_cache: torch.Size([1, 8, 316, 128]) - layer.2.k_cache: torch.Size([1, 8, 316, 128]) - layer.2.v_cache: torch.Size([1, 8, 316, 128]) - layer.3.k_cache: torch.Size([1, 8, 316, 128]) - layer.3.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.k_cache: torch.Size([1, 8, 316, 128]) - layer.4.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.output: torch.Size([1, 316, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02512236 17.77465357 - layer.0.v_cache 0.00000027 0.00029658 - layer.1.k_cache 0.00293329 1.15161896 - layer.1.v_cache 0.00000080 0.00110352 - layer.2.k_cache 0.00117467 0.67628730 - layer.2.v_cache 0.00000111 0.00161445 - layer.3.k_cache 0.00131319 0.74981694 - layer.3.v_cache 0.00000216 0.00269556 - layer.4.k_cache 0.00359311 1.37705028 - layer.4.v_cache 0.00000344 0.00458797 - layer.4.output 0.00015753 0.09554415 - ------------------------------------------------------------------------------------- - TOTAL 0.00248390 1.58013584 - (elements=4,530,176) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4530176 -Total Bytes 518108 -BPFP 0.9149 bits/point -EBPFP 1.8299 equivalent bits/point -MSE 1.580136 ----------------------- -------------------------------------------------------- -Time: 0.964s Load: 0.016s, Pack+Encode: 0.357s, Decode+Unpack: 0.591s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 316, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5801 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample23-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample23-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample24-layer4-item1.zst (28/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample24-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 348, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 348, 128) -Output shape: (1, 348, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.0.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.1.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.1.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.2.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.2.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.3.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.3.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.4.k_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.4.v_cache: torch.Size([1, 8, 348, 128]) -> torch.Size([1, 1, 348, 1024]) - layer.4.output: torch.Size([1, 348, 4096]) -> torch.Size([1, 1, 348, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,980B, BPFP=0.2016 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,336B, BPFP=1.3321 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,372B, BPFP=0.6369 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,940B, BPFP=1.4354 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,980B, BPFP=0.8077 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,184B, BPFP=1.4634 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,452B, BPFP=0.8857 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,672B, BPFP=1.4070 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,756B, BPFP=0.6231 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,340B, BPFP=1.4669 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,072B, BPFP=0.7918 -⌛️ [2/4] FRONTEND: Frontend time: 0.426s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 348, 128]) - layer.0.v_cache: torch.Size([1, 8, 348, 128]) - layer.1.k_cache: torch.Size([1, 8, 348, 128]) - layer.1.v_cache: torch.Size([1, 8, 348, 128]) - layer.2.k_cache: torch.Size([1, 8, 348, 128]) - layer.2.v_cache: torch.Size([1, 8, 348, 128]) - layer.3.k_cache: torch.Size([1, 8, 348, 128]) - layer.3.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.k_cache: torch.Size([1, 8, 348, 128]) - layer.4.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.output: torch.Size([1, 348, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.708s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 348, 128]) - layer.0.v_cache: torch.Size([1, 8, 348, 128]) - layer.1.k_cache: torch.Size([1, 8, 348, 128]) - layer.1.v_cache: torch.Size([1, 8, 348, 128]) - layer.2.k_cache: torch.Size([1, 8, 348, 128]) - layer.2.v_cache: torch.Size([1, 8, 348, 128]) - layer.3.k_cache: torch.Size([1, 8, 348, 128]) - layer.3.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.k_cache: torch.Size([1, 8, 348, 128]) - layer.4.v_cache: torch.Size([1, 8, 348, 128]) - layer.4.output: torch.Size([1, 348, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02368600 18.86896945 - layer.0.v_cache 0.00000027 0.00030973 - layer.1.k_cache 0.00288004 1.23878725 - layer.1.v_cache 0.00000079 0.00115880 - layer.2.k_cache 0.00117070 0.71332366 - layer.2.v_cache 0.00000116 0.00172985 - layer.3.k_cache 0.00131678 0.77039960 - layer.3.v_cache 0.00000214 0.00277093 - layer.4.k_cache 0.00357349 1.42448197 - layer.4.v_cache 0.00000347 0.00468318 - layer.4.output 0.00013697 0.09264973 - ------------------------------------------------------------------------------------- - TOTAL 0.00237019 1.67122952 - (elements=4,988,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4988928 -Total Bytes 598084 -BPFP 0.9591 bits/point -EBPFP 1.9181 equivalent bits/point -MSE 1.671230 ----------------------- -------------------------------------------------------- -Time: 1.152s Load: 0.018s, Pack+Encode: 0.426s, Decode+Unpack: 0.708s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 348, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 348, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6712 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample24-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample24-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample25-layer4-item1.zst (29/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample25-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 345, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 345, 128) -Output shape: (1, 345, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.0.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.1.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.1.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.2.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.2.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.3.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.3.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.4.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.4.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.4.output: torch.Size([1, 345, 4096]) -> torch.Size([1, 1, 345, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,856B, BPFP=0.2005 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,304B, BPFP=1.3429 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,580B, BPFP=0.6245 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,996B, BPFP=1.4492 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,992B, BPFP=0.8377 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,180B, BPFP=1.4760 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,412B, BPFP=0.8925 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,644B, BPFP=1.4412 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,116B, BPFP=0.6140 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,412B, BPFP=1.5039 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 143,876B, BPFP=0.8145 -⌛️ [2/4] FRONTEND: Frontend time: 0.414s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 345, 128]) - layer.0.v_cache: torch.Size([1, 8, 345, 128]) - layer.1.k_cache: torch.Size([1, 8, 345, 128]) - layer.1.v_cache: torch.Size([1, 8, 345, 128]) - layer.2.k_cache: torch.Size([1, 8, 345, 128]) - layer.2.v_cache: torch.Size([1, 8, 345, 128]) - layer.3.k_cache: torch.Size([1, 8, 345, 128]) - layer.3.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.k_cache: torch.Size([1, 8, 345, 128]) - layer.4.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.output: torch.Size([1, 345, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.682s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 345, 128]) - layer.0.v_cache: torch.Size([1, 8, 345, 128]) - layer.1.k_cache: torch.Size([1, 8, 345, 128]) - layer.1.v_cache: torch.Size([1, 8, 345, 128]) - layer.2.k_cache: torch.Size([1, 8, 345, 128]) - layer.2.v_cache: torch.Size([1, 8, 345, 128]) - layer.3.k_cache: torch.Size([1, 8, 345, 128]) - layer.3.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.k_cache: torch.Size([1, 8, 345, 128]) - layer.4.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.output: torch.Size([1, 345, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02357864 18.59168648 - layer.0.v_cache 0.00000026 0.00029150 - layer.1.k_cache 0.00288937 1.24234814 - layer.1.v_cache 0.00000077 0.00113845 - layer.2.k_cache 0.00116899 0.69592710 - layer.2.v_cache 0.00000115 0.00171070 - layer.3.k_cache 0.00131620 0.75448157 - layer.3.v_cache 0.00000216 0.00274599 - layer.4.k_cache 0.00353981 1.40262982 - layer.4.v_cache 0.00000338 0.00472301 - layer.4.output 0.00014567 0.08880972 - ------------------------------------------------------------------------------------- - TOTAL 0.00236310 1.64663726 - (elements=4,945,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4945920 -Total Bytes 602368 -BPFP 0.9743 bits/point -EBPFP 1.9487 equivalent bits/point -MSE 1.646637 ----------------------- -------------------------------------------------------- -Time: 1.113s Load: 0.017s, Pack+Encode: 0.414s, Decode+Unpack: 0.682s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 345, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6466 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample25-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample25-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample26-layer4-item1.zst (30/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample26-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 330, 128) -Output shape: (1, 330, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.output: torch.Size([1, 330, 4096]) -> torch.Size([1, 1, 330, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,968B, BPFP=0.2123 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,400B, BPFP=1.3589 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,504B, BPFP=0.6748 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,208B, BPFP=1.4727 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,196B, BPFP=0.8569 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,256B, BPFP=1.5212 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,540B, BPFP=0.9124 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,744B, BPFP=1.4854 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,792B, BPFP=0.6580 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,896B, BPFP=1.5364 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 153,260B, BPFP=0.9071 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.674s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02434325 20.53059748 - layer.0.v_cache 0.00000027 0.00031039 - layer.1.k_cache 0.00287412 1.20006686 - layer.1.v_cache 0.00000081 0.00114994 - layer.2.k_cache 0.00119273 0.69949738 - layer.2.v_cache 0.00000117 0.00173749 - layer.3.k_cache 0.00130297 0.76743987 - layer.3.v_cache 0.00000221 0.00283947 - layer.4.k_cache 0.00352071 1.38788878 - layer.4.v_cache 0.00000330 0.00481378 - layer.4.output 0.00014370 0.09535748 - ------------------------------------------------------------------------------------- - TOTAL 0.00241545 1.78412653 - (elements=4,730,880) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4730880 -Total Bytes 604764 -BPFP 1.0227 bits/point -EBPFP 2.0453 equivalent bits/point -MSE 1.784127 ----------------------- -------------------------------------------------------- -Time: 1.098s Load: 0.017s, Pack+Encode: 0.406s, Decode+Unpack: 0.674s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7841 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample26-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample26-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample27-layer4-item1.zst (31/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample27-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 334, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 334, 128) -Output shape: (1, 334, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.0.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.1.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.1.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.2.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.2.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.3.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.3.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.4.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.4.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.4.output: torch.Size([1, 334, 4096]) -> torch.Size([1, 1, 334, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,232B, BPFP=0.2159 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,376B, BPFP=1.3421 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,536B, BPFP=0.6441 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,144B, BPFP=1.4536 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,436B, BPFP=0.8289 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,828B, BPFP=1.4930 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,516B, BPFP=0.9009 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,884B, BPFP=1.4475 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,240B, BPFP=0.6372 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,168B, BPFP=1.5009 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,276B, BPFP=0.8437 -⌛️ [2/4] FRONTEND: Frontend time: 0.405s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 334, 128]) - layer.0.v_cache: torch.Size([1, 8, 334, 128]) - layer.1.k_cache: torch.Size([1, 8, 334, 128]) - layer.1.v_cache: torch.Size([1, 8, 334, 128]) - layer.2.k_cache: torch.Size([1, 8, 334, 128]) - layer.2.v_cache: torch.Size([1, 8, 334, 128]) - layer.3.k_cache: torch.Size([1, 8, 334, 128]) - layer.3.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.k_cache: torch.Size([1, 8, 334, 128]) - layer.4.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.output: torch.Size([1, 334, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.695s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 334, 128]) - layer.0.v_cache: torch.Size([1, 8, 334, 128]) - layer.1.k_cache: torch.Size([1, 8, 334, 128]) - layer.1.v_cache: torch.Size([1, 8, 334, 128]) - layer.2.k_cache: torch.Size([1, 8, 334, 128]) - layer.2.v_cache: torch.Size([1, 8, 334, 128]) - layer.3.k_cache: torch.Size([1, 8, 334, 128]) - layer.3.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.k_cache: torch.Size([1, 8, 334, 128]) - layer.4.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.output: torch.Size([1, 334, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02363260 19.38009771 - layer.0.v_cache 0.00000026 0.00030758 - layer.1.k_cache 0.00290221 1.24784440 - layer.1.v_cache 0.00000079 0.00114103 - layer.2.k_cache 0.00116170 0.69596845 - layer.2.v_cache 0.00000115 0.00169371 - layer.3.k_cache 0.00132769 0.77206138 - layer.3.v_cache 0.00000211 0.00271724 - layer.4.k_cache 0.00362300 1.41078707 - layer.4.v_cache 0.00000352 0.00463792 - layer.4.output 0.00014243 0.09590669 - ------------------------------------------------------------------------------------- - TOTAL 0.00237320 1.70720594 - (elements=4,788,224) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4788224 -Total Bytes 591636 -BPFP 0.9885 bits/point -EBPFP 1.9770 equivalent bits/point -MSE 1.707206 ----------------------- -------------------------------------------------------- -Time: 1.117s Load: 0.017s, Pack+Encode: 0.405s, Decode+Unpack: 0.695s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 334, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7072 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample27-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample27-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample28-layer4-item1.zst (32/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample28-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 330, 128) -Output shape: (1, 330, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.output: torch.Size([1, 330, 4096]) -> torch.Size([1, 1, 330, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,896B, BPFP=0.2106 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,368B, BPFP=1.3581 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,492B, BPFP=0.6745 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,008B, BPFP=1.4680 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,128B, BPFP=0.8553 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,052B, BPFP=1.5164 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,696B, BPFP=0.9161 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,656B, BPFP=1.4833 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,192B, BPFP=0.6438 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,820B, BPFP=1.5346 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 149,616B, BPFP=0.8855 -⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02410887 20.50770153 - layer.0.v_cache 0.00000026 0.00030753 - layer.1.k_cache 0.00288475 1.21592093 - layer.1.v_cache 0.00000079 0.00114636 - layer.2.k_cache 0.00119186 0.69645742 - layer.2.v_cache 0.00000115 0.00168004 - layer.3.k_cache 0.00132245 0.76197473 - layer.3.v_cache 0.00000226 0.00281303 - layer.4.k_cache 0.00352035 1.39015512 - layer.4.v_cache 0.00000339 0.00466390 - layer.4.output 0.00014978 0.10036426 - ------------------------------------------------------------------------------------- - TOTAL 0.00240252 1.78459126 - (elements=4,730,880) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4730880 -Total Bytes 599924 -BPFP 1.0145 bits/point -EBPFP 2.0290 equivalent bits/point -MSE 1.784591 ----------------------- -------------------------------------------------------- -Time: 1.094s Load: 0.016s, Pack+Encode: 0.400s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7846 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample28-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample28-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample29-layer4-item1.zst (33/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample29-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 329, 128) -Output shape: (1, 329, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.output: torch.Size([1, 329, 4096]) -> torch.Size([1, 1, 329, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,968B, BPFP=0.2130 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,548B, BPFP=1.3665 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,684B, BPFP=0.6574 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,624B, BPFP=1.4633 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,560B, BPFP=0.8444 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,056B, BPFP=1.4973 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,624B, BPFP=0.9172 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,844B, BPFP=1.4686 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,388B, BPFP=0.6504 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,260B, BPFP=1.5259 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 155,196B, BPFP=0.9213 -⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.676s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02426959 20.15670563 - layer.0.v_cache 0.00000027 0.00030875 - layer.1.k_cache 0.00291730 1.20518925 - layer.1.v_cache 0.00000079 0.00112209 - layer.2.k_cache 0.00116725 0.68803698 - layer.2.v_cache 0.00000113 0.00164357 - layer.3.k_cache 0.00130327 0.76146890 - layer.3.v_cache 0.00000218 0.00272211 - layer.4.k_cache 0.00359671 1.39256301 - layer.4.v_cache 0.00000327 0.00476527 - layer.4.output 0.00014627 0.09427051 - ------------------------------------------------------------------------------------- - TOTAL 0.00241763 1.75654340 - (elements=4,716,544) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4716544 -Total Bytes 601752 -BPFP 1.0207 bits/point -EBPFP 2.0413 equivalent bits/point -MSE 1.756543 ----------------------- -------------------------------------------------------- -Time: 1.094s Load: 0.017s, Pack+Encode: 0.400s, Decode+Unpack: 0.676s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7565 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample29-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample29-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample3-layer4-item1.zst (34/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample3-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 365, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 365, 128) -Output shape: (1, 365, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.0.v_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.1.k_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.1.v_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.2.k_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.2.v_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.3.k_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.3.v_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.4.k_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.4.v_cache: torch.Size([1, 8, 365, 128]) -> torch.Size([1, 1, 365, 1024]) - layer.4.output: torch.Size([1, 365, 4096]) -> torch.Size([1, 1, 365, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,316B, BPFP=0.1994 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,472B, BPFP=1.2729 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,528B, BPFP=0.6106 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,728B, BPFP=1.3854 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,616B, BPFP=0.8051 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,764B, BPFP=1.4076 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,988B, BPFP=0.8559 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,876B, BPFP=1.3672 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,236B, BPFP=0.6044 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,420B, BPFP=1.4217 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 150,148B, BPFP=0.8034 -⌛️ [2/4] FRONTEND: Frontend time: 0.403s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 365, 128]) - layer.0.v_cache: torch.Size([1, 8, 365, 128]) - layer.1.k_cache: torch.Size([1, 8, 365, 128]) - layer.1.v_cache: torch.Size([1, 8, 365, 128]) - layer.2.k_cache: torch.Size([1, 8, 365, 128]) - layer.2.v_cache: torch.Size([1, 8, 365, 128]) - layer.3.k_cache: torch.Size([1, 8, 365, 128]) - layer.3.v_cache: torch.Size([1, 8, 365, 128]) - layer.4.k_cache: torch.Size([1, 8, 365, 128]) - layer.4.v_cache: torch.Size([1, 8, 365, 128]) - layer.4.output: torch.Size([1, 365, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.675s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 365, 128]) - layer.0.v_cache: torch.Size([1, 8, 365, 128]) - layer.1.k_cache: torch.Size([1, 8, 365, 128]) - layer.1.v_cache: torch.Size([1, 8, 365, 128]) - layer.2.k_cache: torch.Size([1, 8, 365, 128]) - layer.2.v_cache: torch.Size([1, 8, 365, 128]) - layer.3.k_cache: torch.Size([1, 8, 365, 128]) - layer.3.v_cache: torch.Size([1, 8, 365, 128]) - layer.4.k_cache: torch.Size([1, 8, 365, 128]) - layer.4.v_cache: torch.Size([1, 8, 365, 128]) - layer.4.output: torch.Size([1, 365, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02487199 18.90691754 - layer.0.v_cache 0.00000027 0.00030835 - layer.1.k_cache 0.00290279 1.22718355 - layer.1.v_cache 0.00000079 0.00116156 - layer.2.k_cache 0.00116885 0.70489285 - layer.2.v_cache 0.00000114 0.00173097 - layer.3.k_cache 0.00131452 0.75018369 - layer.3.v_cache 0.00000216 0.00279188 - layer.4.k_cache 0.00360833 1.39179704 - layer.4.v_cache 0.00000353 0.00466341 - layer.4.output 0.00014951 0.08876396 - ------------------------------------------------------------------------------------- - TOTAL 0.00246232 1.66762048 - (elements=5,232,640) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5232640 -Total Bytes 614092 -BPFP 0.9389 bits/point -EBPFP 1.8777 equivalent bits/point -MSE 1.667620 ----------------------- -------------------------------------------------------- -Time: 1.097s Load: 0.019s, Pack+Encode: 0.403s, Decode+Unpack: 0.675s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 365, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 365, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6676 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample3-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample3-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample30-layer4-item1.zst (35/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample30-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 335, 128) -Output shape: (1, 335, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.output: torch.Size([1, 335, 4096]) -> torch.Size([1, 1, 335, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,180B, BPFP=0.2141 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,116B, BPFP=1.3553 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,604B, BPFP=0.6671 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,928B, BPFP=1.4675 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,332B, BPFP=0.8473 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,168B, BPFP=1.4965 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,980B, BPFP=0.9090 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,420B, BPFP=1.4557 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,128B, BPFP=0.6326 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,596B, BPFP=1.5064 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,440B, BPFP=0.8479 -⌛️ [2/4] FRONTEND: Frontend time: 0.404s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.675s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02462482 18.64884999 - layer.0.v_cache 0.00000027 0.00030519 - layer.1.k_cache 0.00287179 1.19498564 - layer.1.v_cache 0.00000079 0.00116260 - layer.2.k_cache 0.00118919 0.70233883 - layer.2.v_cache 0.00000114 0.00169937 - layer.3.k_cache 0.00130718 0.75587104 - layer.3.v_cache 0.00000215 0.00272855 - layer.4.k_cache 0.00355403 1.37306582 - layer.4.v_cache 0.00000335 0.00458994 - layer.4.output 0.00013864 0.09078762 - ------------------------------------------------------------------------------------- - TOTAL 0.00243638 1.64633910 - (elements=4,802,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4802560 -Total Bytes 597892 -BPFP 0.9960 bits/point -EBPFP 1.9919 equivalent bits/point -MSE 1.646339 ----------------------- -------------------------------------------------------- -Time: 1.096s Load: 0.017s, Pack+Encode: 0.404s, Decode+Unpack: 0.675s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6463 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample30-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample30-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample31-layer4-item1.zst (36/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample31-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 335, 128) -Output shape: (1, 335, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.output: torch.Size([1, 335, 4096]) -> torch.Size([1, 1, 335, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,128B, BPFP=0.2129 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,684B, BPFP=1.3452 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,016B, BPFP=0.6534 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,548B, BPFP=1.4587 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,324B, BPFP=0.8471 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,096B, BPFP=1.4948 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,820B, BPFP=0.9053 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,436B, BPFP=1.4561 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,712B, BPFP=0.6229 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,520B, BPFP=1.5047 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,268B, BPFP=0.8411 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.680s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02394044 18.11890013 - layer.0.v_cache 0.00000026 0.00029745 - layer.1.k_cache 0.00289851 1.20097902 - layer.1.v_cache 0.00000079 0.00112776 - layer.2.k_cache 0.00115963 0.70203252 - layer.2.v_cache 0.00000116 0.00170914 - layer.3.k_cache 0.00131431 0.76181823 - layer.3.v_cache 0.00000219 0.00279110 - layer.4.k_cache 0.00369786 1.40622641 - layer.4.v_cache 0.00000335 0.00457669 - layer.4.output 0.00014933 0.09302047 - ------------------------------------------------------------------------------------- - TOTAL 0.00240113 1.61232431 - (elements=4,802,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4802560 -Total Bytes 594552 -BPFP 0.9904 bits/point -EBPFP 1.9808 equivalent bits/point -MSE 1.612324 ----------------------- -------------------------------------------------------- -Time: 1.099s Load: 0.017s, Pack+Encode: 0.402s, Decode+Unpack: 0.680s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6123 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample31-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample31-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample32-layer4-item1.zst (37/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample32-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 339, 128) -Output shape: (1, 339, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.output: torch.Size([1, 339, 4096]) -> torch.Size([1, 1, 339, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,104B, BPFP=0.2098 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,524B, BPFP=1.3487 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,548B, BPFP=0.6579 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,228B, BPFP=1.4571 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,660B, BPFP=0.8449 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,868B, BPFP=1.4949 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,068B, BPFP=0.9004 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,236B, BPFP=1.4573 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,392B, BPFP=0.6313 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,168B, BPFP=1.5018 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 149,092B, BPFP=0.8590 -⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.685s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02382689 19.39014680 - layer.0.v_cache 0.00000027 0.00031144 - layer.1.k_cache 0.00292607 1.24838419 - layer.1.v_cache 0.00000081 0.00113807 - layer.2.k_cache 0.00125336 0.71045543 - layer.2.v_cache 0.00000116 0.00172779 - layer.3.k_cache 0.00131849 0.75782978 - layer.3.v_cache 0.00000220 0.00280058 - layer.4.k_cache 0.00356446 1.38938584 - layer.4.v_cache 0.00000343 0.00472379 - layer.4.output 0.00016600 0.09299767 - ------------------------------------------------------------------------------------- - TOTAL 0.00239722 1.70563531 - (elements=4,859,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4859904 -Total Bytes 604888 -BPFP 0.9957 bits/point -EBPFP 1.9914 equivalent bits/point -MSE 1.705635 ----------------------- -------------------------------------------------------- -Time: 1.103s Load: 0.018s, Pack+Encode: 0.400s, Decode+Unpack: 0.685s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7056 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample32-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample32-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample33-layer4-item1.zst (38/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample33-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 362, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 362, 128) -Output shape: (1, 362, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.0.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.1.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.1.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.2.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.2.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.3.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.3.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.4.k_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.4.v_cache: torch.Size([1, 8, 362, 128]) -> torch.Size([1, 1, 362, 1024]) - layer.4.output: torch.Size([1, 362, 4096]) -> torch.Size([1, 1, 362, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,088B, BPFP=0.1961 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,652B, BPFP=1.2874 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,140B, BPFP=0.6289 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 65,028B, BPFP=1.4034 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 38,252B, BPFP=0.8255 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,412B, BPFP=1.4117 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,260B, BPFP=0.8689 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,760B, BPFP=1.3760 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,836B, BPFP=0.6223 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,224B, BPFP=1.4292 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,216B, BPFP=0.7835 -⌛️ [2/4] FRONTEND: Frontend time: 0.412s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 362, 128]) - layer.0.v_cache: torch.Size([1, 8, 362, 128]) - layer.1.k_cache: torch.Size([1, 8, 362, 128]) - layer.1.v_cache: torch.Size([1, 8, 362, 128]) - layer.2.k_cache: torch.Size([1, 8, 362, 128]) - layer.2.v_cache: torch.Size([1, 8, 362, 128]) - layer.3.k_cache: torch.Size([1, 8, 362, 128]) - layer.3.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.k_cache: torch.Size([1, 8, 362, 128]) - layer.4.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.output: torch.Size([1, 362, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.679s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 362, 128]) - layer.0.v_cache: torch.Size([1, 8, 362, 128]) - layer.1.k_cache: torch.Size([1, 8, 362, 128]) - layer.1.v_cache: torch.Size([1, 8, 362, 128]) - layer.2.k_cache: torch.Size([1, 8, 362, 128]) - layer.2.v_cache: torch.Size([1, 8, 362, 128]) - layer.3.k_cache: torch.Size([1, 8, 362, 128]) - layer.3.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.k_cache: torch.Size([1, 8, 362, 128]) - layer.4.v_cache: torch.Size([1, 8, 362, 128]) - layer.4.output: torch.Size([1, 362, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02338883 19.82018976 - layer.0.v_cache 0.00000028 0.00031429 - layer.1.k_cache 0.00287711 1.23861593 - layer.1.v_cache 0.00000077 0.00115124 - layer.2.k_cache 0.00119138 0.70406021 - layer.2.v_cache 0.00000112 0.00169233 - layer.3.k_cache 0.00131510 0.75931418 - layer.3.v_cache 0.00000211 0.00271996 - layer.4.k_cache 0.00352384 1.38496635 - layer.4.v_cache 0.00000335 0.00464935 - layer.4.output 0.00013081 0.08912823 - ------------------------------------------------------------------------------------- - TOTAL 0.00234480 1.73387047 - (elements=5,189,632) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5189632 -Total Bytes 610868 -BPFP 0.9417 bits/point -EBPFP 1.8833 equivalent bits/point -MSE 1.733870 ----------------------- -------------------------------------------------------- -Time: 1.109s Load: 0.018s, Pack+Encode: 0.412s, Decode+Unpack: 0.679s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 362, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 362, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7339 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample33-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample33-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample34-layer4-item1.zst (39/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample34-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 318, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 318, 128) -Output shape: (1, 318, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.0.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.1.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.1.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.2.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.2.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.3.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.3.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.4.k_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.4.v_cache: torch.Size([1, 8, 318, 128]) -> torch.Size([1, 1, 318, 1024]) - layer.4.output: torch.Size([1, 318, 4096]) -> torch.Size([1, 1, 318, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,860B, BPFP=0.2177 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,620B, BPFP=1.2436 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,668B, BPFP=0.6060 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 55,156B, BPFP=1.3551 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,268B, BPFP=0.7927 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,584B, BPFP=1.3656 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,500B, BPFP=0.8476 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 54,420B, BPFP=1.3370 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,480B, BPFP=0.6014 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,188B, BPFP=1.3804 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 131,356B, BPFP=0.8068 -⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 318, 128]) - layer.0.v_cache: torch.Size([1, 8, 318, 128]) - layer.1.k_cache: torch.Size([1, 8, 318, 128]) - layer.1.v_cache: torch.Size([1, 8, 318, 128]) - layer.2.k_cache: torch.Size([1, 8, 318, 128]) - layer.2.v_cache: torch.Size([1, 8, 318, 128]) - layer.3.k_cache: torch.Size([1, 8, 318, 128]) - layer.3.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.k_cache: torch.Size([1, 8, 318, 128]) - layer.4.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.output: torch.Size([1, 318, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.573s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 318, 128]) - layer.0.v_cache: torch.Size([1, 8, 318, 128]) - layer.1.k_cache: torch.Size([1, 8, 318, 128]) - layer.1.v_cache: torch.Size([1, 8, 318, 128]) - layer.2.k_cache: torch.Size([1, 8, 318, 128]) - layer.2.v_cache: torch.Size([1, 8, 318, 128]) - layer.3.k_cache: torch.Size([1, 8, 318, 128]) - layer.3.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.k_cache: torch.Size([1, 8, 318, 128]) - layer.4.v_cache: torch.Size([1, 8, 318, 128]) - layer.4.output: torch.Size([1, 318, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02477463 17.39219119 - layer.0.v_cache 0.00000027 0.00030377 - layer.1.k_cache 0.00289381 1.19327924 - layer.1.v_cache 0.00000083 0.00115674 - layer.2.k_cache 0.00119129 0.69875091 - layer.2.v_cache 0.00000117 0.00172799 - layer.3.k_cache 0.00131631 0.77207721 - layer.3.v_cache 0.00000257 0.00284440 - layer.4.k_cache 0.00356231 1.39159542 - layer.4.v_cache 0.00000346 0.00476015 - layer.4.output 0.00015646 0.09821607 - ------------------------------------------------------------------------------------- - TOTAL 0.00245518 1.56082509 - (elements=4,558,848) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4558848 -Total Bytes 528100 -BPFP 0.9267 bits/point -EBPFP 1.8535 equivalent bits/point -MSE 1.560825 ----------------------- -------------------------------------------------------- -Time: 0.944s Load: 0.017s, Pack+Encode: 0.354s, Decode+Unpack: 0.573s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 318, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 318, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5608 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample34-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample34-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample35-layer4-item1.zst (40/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample35-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 342, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 342, 128) -Output shape: (1, 342, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.output: torch.Size([1, 342, 4096]) -> torch.Size([1, 1, 342, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,960B, BPFP=0.2047 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,132B, BPFP=1.3508 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,928B, BPFP=0.6380 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,076B, BPFP=1.4637 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,456B, BPFP=0.8328 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,388B, BPFP=1.4937 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,336B, BPFP=0.8986 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,460B, BPFP=1.4497 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,124B, BPFP=0.6196 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,896B, BPFP=1.5053 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,732B, BPFP=0.8094 -⌛️ [2/4] FRONTEND: Frontend time: 0.415s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.output: torch.Size([1, 342, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.690s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.output: torch.Size([1, 342, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02487079 18.92318279 - layer.0.v_cache 0.00000026 0.00030075 - layer.1.k_cache 0.00288341 1.21989682 - layer.1.v_cache 0.00000082 0.00114753 - layer.2.k_cache 0.00117177 0.69485465 - layer.2.v_cache 0.00000117 0.00168963 - layer.3.k_cache 0.00130255 0.75478832 - layer.3.v_cache 0.00000215 0.00273713 - layer.4.k_cache 0.00351771 1.33971757 - layer.4.v_cache 0.00000337 0.00463199 - layer.4.output 0.00013439 0.08789179 - ------------------------------------------------------------------------------------- - TOTAL 0.00244940 1.66389388 - (elements=4,902,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4902912 -Total Bytes 599488 -BPFP 0.9782 bits/point -EBPFP 1.9563 equivalent bits/point -MSE 1.663894 ----------------------- -------------------------------------------------------- -Time: 1.123s Load: 0.018s, Pack+Encode: 0.415s, Decode+Unpack: 0.690s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 342, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6639 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample35-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample35-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample36-layer4-item1.zst (41/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample36-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 343, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 343, 128) -Output shape: (1, 343, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.output: torch.Size([1, 343, 4096]) -> torch.Size([1, 1, 343, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,028B, BPFP=0.2056 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,472B, BPFP=1.3546 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,180B, BPFP=0.6419 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,028B, BPFP=1.4584 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,816B, BPFP=0.8386 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,276B, BPFP=1.4868 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,288B, BPFP=0.8949 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,448B, BPFP=1.4452 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,936B, BPFP=0.6135 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,232B, BPFP=1.5086 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 143,928B, BPFP=0.8196 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.output: torch.Size([1, 343, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.720s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.output: torch.Size([1, 343, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02787456 19.44810268 - layer.0.v_cache 0.00000027 0.00030442 - layer.1.k_cache 0.00288980 1.20235592 - layer.1.v_cache 0.00000078 0.00116477 - layer.2.k_cache 0.00118800 0.69563952 - layer.2.v_cache 0.00000116 0.00169478 - layer.3.k_cache 0.00131625 0.75796838 - layer.3.v_cache 0.00000215 0.00278003 - layer.4.k_cache 0.00354145 1.37284606 - layer.4.v_cache 0.00000355 0.00470977 - layer.4.output 0.00016614 0.09548276 - ------------------------------------------------------------------------------------- - TOTAL 0.00267732 1.70496410 - (elements=4,917,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4917248 -Total Bytes 602632 -BPFP 0.9804 bits/point -EBPFP 1.9609 equivalent bits/point -MSE 1.704964 ----------------------- -------------------------------------------------------- -Time: 1.144s Load: 0.018s, Pack+Encode: 0.406s, Decode+Unpack: 0.720s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 343, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7050 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample36-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample36-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample37-layer4-item1.zst (42/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample37-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 333, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 333, 128) -Output shape: (1, 333, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.output: torch.Size([1, 333, 4096]) -> torch.Size([1, 1, 333, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,092B, BPFP=0.2133 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,732B, BPFP=1.3544 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,708B, BPFP=0.6501 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,672B, BPFP=1.4703 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,708B, BPFP=0.8377 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,448B, BPFP=1.5120 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,188B, BPFP=0.8959 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,444B, BPFP=1.4650 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,900B, BPFP=0.6311 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,992B, BPFP=1.5248 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 140,584B, BPFP=0.8246 -⌛️ [2/4] FRONTEND: Frontend time: 0.401s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.output: torch.Size([1, 333, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.679s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.output: torch.Size([1, 333, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02325115 19.37592817 - layer.0.v_cache 0.00000026 0.00030548 - layer.1.k_cache 0.00292822 1.18331653 - layer.1.v_cache 0.00000083 0.00115154 - layer.2.k_cache 0.00115284 0.70288975 - layer.2.v_cache 0.00000118 0.00168738 - layer.3.k_cache 0.00132958 0.75630133 - layer.3.v_cache 0.00000212 0.00271718 - layer.4.k_cache 0.00351398 1.40776452 - layer.4.v_cache 0.00000364 0.00465052 - layer.4.output 0.00014544 0.09337822 - ------------------------------------------------------------------------------------- - TOTAL 0.00234040 1.70073038 - (elements=4,773,888) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4773888 -Total Bytes 590468 -BPFP 0.9895 bits/point -EBPFP 1.9790 equivalent bits/point -MSE 1.700730 ----------------------- -------------------------------------------------------- -Time: 1.096s Load: 0.017s, Pack+Encode: 0.401s, Decode+Unpack: 0.679s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 333, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7007 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample37-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample37-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample38-layer4-item1.zst (43/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample38-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 356, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 356, 128) -Output shape: (1, 356, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.0.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.1.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.1.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.2.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.2.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.3.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.3.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.4.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.4.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.4.output: torch.Size([1, 356, 4096]) -> torch.Size([1, 1, 356, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,936B, BPFP=0.1961 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,540B, BPFP=1.3066 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,800B, BPFP=0.6320 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,844B, BPFP=1.4011 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,156B, BPFP=0.8154 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,604B, BPFP=1.4177 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,816B, BPFP=0.8738 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,936B, BPFP=1.3811 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,280B, BPFP=0.5987 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,716B, BPFP=1.4422 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 146,928B, BPFP=0.8061 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 356, 128]) - layer.0.v_cache: torch.Size([1, 8, 356, 128]) - layer.1.k_cache: torch.Size([1, 8, 356, 128]) - layer.1.v_cache: torch.Size([1, 8, 356, 128]) - layer.2.k_cache: torch.Size([1, 8, 356, 128]) - layer.2.v_cache: torch.Size([1, 8, 356, 128]) - layer.3.k_cache: torch.Size([1, 8, 356, 128]) - layer.3.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.k_cache: torch.Size([1, 8, 356, 128]) - layer.4.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.output: torch.Size([1, 356, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.679s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 356, 128]) - layer.0.v_cache: torch.Size([1, 8, 356, 128]) - layer.1.k_cache: torch.Size([1, 8, 356, 128]) - layer.1.v_cache: torch.Size([1, 8, 356, 128]) - layer.2.k_cache: torch.Size([1, 8, 356, 128]) - layer.2.v_cache: torch.Size([1, 8, 356, 128]) - layer.3.k_cache: torch.Size([1, 8, 356, 128]) - layer.3.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.k_cache: torch.Size([1, 8, 356, 128]) - layer.4.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.output: torch.Size([1, 356, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02485108 18.89266728 - layer.0.v_cache 0.00000027 0.00030630 - layer.1.k_cache 0.00289550 1.20595499 - layer.1.v_cache 0.00000079 0.00114223 - layer.2.k_cache 0.00117641 0.71377653 - layer.2.v_cache 0.00000113 0.00169547 - layer.3.k_cache 0.00131994 0.74984887 - layer.3.v_cache 0.00000217 0.00277841 - layer.4.k_cache 0.00362269 1.43242988 - layer.4.v_cache 0.00000343 0.00471582 - layer.4.output 0.00016041 0.09348144 - ------------------------------------------------------------------------------------- - TOTAL 0.00246536 1.66994582 - (elements=5,103,616) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5103616 -Total Bytes 605556 -BPFP 0.9492 bits/point -EBPFP 1.8984 equivalent bits/point -MSE 1.669946 ----------------------- -------------------------------------------------------- -Time: 1.104s Load: 0.019s, Pack+Encode: 0.406s, Decode+Unpack: 0.679s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 356, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6699 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample38-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample38-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample39-layer4-item1.zst (44/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample39-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 331, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 331, 128) -Output shape: (1, 331, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.0.v_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.1.k_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.1.v_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.2.k_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.2.v_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.3.k_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.3.v_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.4.k_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.4.v_cache: torch.Size([1, 8, 331, 128]) -> torch.Size([1, 1, 331, 1024]) - layer.4.output: torch.Size([1, 331, 4096]) -> torch.Size([1, 1, 331, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,132B, BPFP=0.2155 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,032B, BPFP=1.3461 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,920B, BPFP=0.6590 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,576B, BPFP=1.4770 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,096B, BPFP=0.8520 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,336B, BPFP=1.5185 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,488B, BPFP=0.9084 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,604B, BPFP=1.4776 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,356B, BPFP=0.6457 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,632B, BPFP=1.5255 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 146,200B, BPFP=0.8627 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 331, 128]) - layer.0.v_cache: torch.Size([1, 8, 331, 128]) - layer.1.k_cache: torch.Size([1, 8, 331, 128]) - layer.1.v_cache: torch.Size([1, 8, 331, 128]) - layer.2.k_cache: torch.Size([1, 8, 331, 128]) - layer.2.v_cache: torch.Size([1, 8, 331, 128]) - layer.3.k_cache: torch.Size([1, 8, 331, 128]) - layer.3.v_cache: torch.Size([1, 8, 331, 128]) - layer.4.k_cache: torch.Size([1, 8, 331, 128]) - layer.4.v_cache: torch.Size([1, 8, 331, 128]) - layer.4.output: torch.Size([1, 331, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.710s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 331, 128]) - layer.0.v_cache: torch.Size([1, 8, 331, 128]) - layer.1.k_cache: torch.Size([1, 8, 331, 128]) - layer.1.v_cache: torch.Size([1, 8, 331, 128]) - layer.2.k_cache: torch.Size([1, 8, 331, 128]) - layer.2.v_cache: torch.Size([1, 8, 331, 128]) - layer.3.k_cache: torch.Size([1, 8, 331, 128]) - layer.3.v_cache: torch.Size([1, 8, 331, 128]) - layer.4.k_cache: torch.Size([1, 8, 331, 128]) - layer.4.v_cache: torch.Size([1, 8, 331, 128]) - layer.4.output: torch.Size([1, 331, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02326727 19.96664936 - layer.0.v_cache 0.00000026 0.00029993 - layer.1.k_cache 0.00291335 1.17653635 - layer.1.v_cache 0.00000080 0.00115586 - layer.2.k_cache 0.00116981 0.71233301 - layer.2.v_cache 0.00000115 0.00168634 - layer.3.k_cache 0.00132745 0.77030143 - layer.3.v_cache 0.00000216 0.00268546 - layer.4.k_cache 0.00359183 1.38591036 - layer.4.v_cache 0.00000354 0.00460753 - layer.4.output 0.00014916 0.09349244 - ------------------------------------------------------------------------------------- - TOTAL 0.00234816 1.74258110 - (elements=4,745,216) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4745216 -Total Bytes 596372 -BPFP 1.0054 bits/point -EBPFP 2.0109 equivalent bits/point -MSE 1.742581 ----------------------- -------------------------------------------------------- -Time: 1.136s Load: 0.017s, Pack+Encode: 0.408s, Decode+Unpack: 0.710s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 331, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 331, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7426 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample39-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample39-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample4-layer4-item1.zst (45/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample4-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 345, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 345, 128) -Output shape: (1, 345, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.0.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.1.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.1.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.2.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.2.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.3.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.3.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.4.k_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.4.v_cache: torch.Size([1, 8, 345, 128]) -> torch.Size([1, 1, 345, 1024]) - layer.4.output: torch.Size([1, 345, 4096]) -> torch.Size([1, 1, 345, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,992B, BPFP=0.2036 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,760B, BPFP=1.3306 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,604B, BPFP=0.6251 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,060B, BPFP=1.4280 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,832B, BPFP=0.8114 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,872B, BPFP=1.4690 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,948B, BPFP=0.8820 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,952B, BPFP=1.4255 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,280B, BPFP=0.5951 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,436B, BPFP=1.4818 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,412B, BPFP=0.8402 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 345, 128]) - layer.0.v_cache: torch.Size([1, 8, 345, 128]) - layer.1.k_cache: torch.Size([1, 8, 345, 128]) - layer.1.v_cache: torch.Size([1, 8, 345, 128]) - layer.2.k_cache: torch.Size([1, 8, 345, 128]) - layer.2.v_cache: torch.Size([1, 8, 345, 128]) - layer.3.k_cache: torch.Size([1, 8, 345, 128]) - layer.3.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.k_cache: torch.Size([1, 8, 345, 128]) - layer.4.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.output: torch.Size([1, 345, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.683s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 345, 128]) - layer.0.v_cache: torch.Size([1, 8, 345, 128]) - layer.1.k_cache: torch.Size([1, 8, 345, 128]) - layer.1.v_cache: torch.Size([1, 8, 345, 128]) - layer.2.k_cache: torch.Size([1, 8, 345, 128]) - layer.2.v_cache: torch.Size([1, 8, 345, 128]) - layer.3.k_cache: torch.Size([1, 8, 345, 128]) - layer.3.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.k_cache: torch.Size([1, 8, 345, 128]) - layer.4.v_cache: torch.Size([1, 8, 345, 128]) - layer.4.output: torch.Size([1, 345, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02346668 19.24468552 - layer.0.v_cache 0.00000026 0.00029911 - layer.1.k_cache 0.00293293 1.30107112 - layer.1.v_cache 0.00000076 0.00106119 - layer.2.k_cache 0.00115776 0.68525492 - layer.2.v_cache 0.00000114 0.00162681 - layer.3.k_cache 0.00132774 0.76146585 - layer.3.v_cache 0.00000222 0.00272397 - layer.4.k_cache 0.00355356 1.40723691 - layer.4.v_cache 0.00000324 0.00454647 - layer.4.output 0.00016540 0.09727256 - ------------------------------------------------------------------------------------- - TOTAL 0.00236485 1.69993301 - (elements=4,945,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4945920 -Total Bytes 601148 -BPFP 0.9724 bits/point -EBPFP 1.9447 equivalent bits/point -MSE 1.699933 ----------------------- -------------------------------------------------------- -Time: 1.110s Load: 0.018s, Pack+Encode: 0.408s, Decode+Unpack: 0.683s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 345, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 345, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6999 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample4-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample4-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample40-layer4-item1.zst (46/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample40-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 344, 128) -Output shape: (1, 344, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.output: torch.Size([1, 344, 4096]) -> torch.Size([1, 1, 344, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,264B, BPFP=0.2104 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,452B, BPFP=1.3502 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,664B, BPFP=0.6283 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,824B, BPFP=1.4495 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,460B, BPFP=0.8280 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,296B, BPFP=1.4829 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,428B, BPFP=0.8954 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,500B, BPFP=1.4421 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,240B, BPFP=0.6186 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,272B, BPFP=1.5051 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,060B, BPFP=0.8179 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.689s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02380104 18.86373547 - layer.0.v_cache 0.00000026 0.00029921 - layer.1.k_cache 0.00287155 1.22348306 - layer.1.v_cache 0.00000078 0.00111825 - layer.2.k_cache 0.00117748 0.69163544 - layer.2.v_cache 0.00000114 0.00163656 - layer.3.k_cache 0.00132414 0.74347354 - layer.3.v_cache 0.00000215 0.00267666 - layer.4.k_cache 0.00352424 1.38555376 - layer.4.v_cache 0.00000349 0.00456866 - layer.4.output 0.00015726 0.08749114 - ------------------------------------------------------------------------------------- - TOTAL 0.00238109 1.66201037 - (elements=4,931,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4931584 -Total Bytes 602460 -BPFP 0.9773 bits/point -EBPFP 1.9546 equivalent bits/point -MSE 1.662010 ----------------------- -------------------------------------------------------- -Time: 1.115s Load: 0.018s, Pack+Encode: 0.408s, Decode+Unpack: 0.689s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6620 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample40-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample40-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample41-layer4-item1.zst (47/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample41-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 339, 128) -Output shape: (1, 339, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.output: torch.Size([1, 339, 4096]) -> torch.Size([1, 1, 339, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,024B, BPFP=0.2080 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,500B, BPFP=1.3482 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,064B, BPFP=0.6468 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,632B, BPFP=1.4664 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,272B, BPFP=0.8359 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,856B, BPFP=1.4947 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,948B, BPFP=0.8976 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,936B, BPFP=1.4504 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,584B, BPFP=0.6357 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,316B, BPFP=1.5053 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,668B, BPFP=0.8162 -⌛️ [2/4] FRONTEND: Frontend time: 0.411s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.685s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02519707 18.37417179 - layer.0.v_cache 0.00000026 0.00030930 - layer.1.k_cache 0.00285639 1.21855438 - layer.1.v_cache 0.00000078 0.00116129 - layer.2.k_cache 0.00115969 0.70158634 - layer.2.v_cache 0.00000117 0.00168971 - layer.3.k_cache 0.00131371 0.76799088 - layer.3.v_cache 0.00000219 0.00273157 - layer.4.k_cache 0.00358480 1.40859634 - layer.4.v_cache 0.00000371 0.00467180 - layer.4.output 0.00015316 0.09429749 - ------------------------------------------------------------------------------------- - TOTAL 0.00248089 1.63276096 - (elements=4,859,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4859904 -Total Bytes 596800 -BPFP 0.9824 bits/point -EBPFP 1.9648 equivalent bits/point -MSE 1.632761 ----------------------- -------------------------------------------------------- -Time: 1.114s Load: 0.018s, Pack+Encode: 0.411s, Decode+Unpack: 0.685s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6328 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample41-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample41-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample42-layer4-item1.zst (48/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample42-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 316, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 316, 128) -Output shape: (1, 316, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.0.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.1.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.1.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.2.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.2.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.3.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.3.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.4.k_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.4.v_cache: torch.Size([1, 8, 316, 128]) -> torch.Size([1, 1, 316, 1024]) - layer.4.output: torch.Size([1, 316, 4096]) -> torch.Size([1, 1, 316, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,400B, BPFP=0.2077 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,724B, BPFP=1.2541 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,404B, BPFP=0.6033 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 55,180B, BPFP=1.3642 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,244B, BPFP=0.7972 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,784B, BPFP=1.3792 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,396B, BPFP=0.8504 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 54,544B, BPFP=1.3485 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,936B, BPFP=0.5918 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,640B, BPFP=1.4003 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 129,428B, BPFP=0.8000 -⌛️ [2/4] FRONTEND: Frontend time: 0.351s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 316, 128]) - layer.0.v_cache: torch.Size([1, 8, 316, 128]) - layer.1.k_cache: torch.Size([1, 8, 316, 128]) - layer.1.v_cache: torch.Size([1, 8, 316, 128]) - layer.2.k_cache: torch.Size([1, 8, 316, 128]) - layer.2.v_cache: torch.Size([1, 8, 316, 128]) - layer.3.k_cache: torch.Size([1, 8, 316, 128]) - layer.3.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.k_cache: torch.Size([1, 8, 316, 128]) - layer.4.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.output: torch.Size([1, 316, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.574s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 316, 128]) - layer.0.v_cache: torch.Size([1, 8, 316, 128]) - layer.1.k_cache: torch.Size([1, 8, 316, 128]) - layer.1.v_cache: torch.Size([1, 8, 316, 128]) - layer.2.k_cache: torch.Size([1, 8, 316, 128]) - layer.2.v_cache: torch.Size([1, 8, 316, 128]) - layer.3.k_cache: torch.Size([1, 8, 316, 128]) - layer.3.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.k_cache: torch.Size([1, 8, 316, 128]) - layer.4.v_cache: torch.Size([1, 8, 316, 128]) - layer.4.output: torch.Size([1, 316, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02408213 18.57629085 - layer.0.v_cache 0.00000026 0.00030093 - layer.1.k_cache 0.00292725 1.14768673 - layer.1.v_cache 0.00000082 0.00112106 - layer.2.k_cache 0.00117417 0.69076601 - layer.2.v_cache 0.00000116 0.00168956 - layer.3.k_cache 0.00131215 0.75612481 - layer.3.v_cache 0.00000217 0.00278534 - layer.4.k_cache 0.00352248 1.39102868 - layer.4.v_cache 0.00000354 0.00470976 - layer.4.output 0.00015546 0.09864024 - ------------------------------------------------------------------------------------- - TOTAL 0.00240343 1.64050462 - (elements=4,530,176) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4530176 -Total Bytes 525680 -BPFP 0.9283 bits/point -EBPFP 1.8566 equivalent bits/point -MSE 1.640505 ----------------------- -------------------------------------------------------- -Time: 0.942s Load: 0.017s, Pack+Encode: 0.351s, Decode+Unpack: 0.574s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 316, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 316, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6405 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample42-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample42-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample43-layer4-item1.zst (49/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample43-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 343, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 343, 128) -Output shape: (1, 343, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) -> torch.Size([1, 1, 343, 1024]) - layer.4.output: torch.Size([1, 343, 4096]) -> torch.Size([1, 1, 343, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,232B, BPFP=0.2103 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,348B, BPFP=1.3518 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,336B, BPFP=0.6454 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,584B, BPFP=1.4710 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,968B, BPFP=0.8420 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,380B, BPFP=1.4892 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,692B, BPFP=0.9041 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,548B, BPFP=1.4474 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,548B, BPFP=0.6275 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,884B, BPFP=1.5006 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 139,272B, BPFP=0.7930 -⌛️ [2/4] FRONTEND: Frontend time: 0.399s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.output: torch.Size([1, 343, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.675s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 343, 128]) - layer.0.v_cache: torch.Size([1, 8, 343, 128]) - layer.1.k_cache: torch.Size([1, 8, 343, 128]) - layer.1.v_cache: torch.Size([1, 8, 343, 128]) - layer.2.k_cache: torch.Size([1, 8, 343, 128]) - layer.2.v_cache: torch.Size([1, 8, 343, 128]) - layer.3.k_cache: torch.Size([1, 8, 343, 128]) - layer.3.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.k_cache: torch.Size([1, 8, 343, 128]) - layer.4.v_cache: torch.Size([1, 8, 343, 128]) - layer.4.output: torch.Size([1, 343, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02365684 18.49447516 - layer.0.v_cache 0.00000027 0.00030705 - layer.1.k_cache 0.00286131 1.21147125 - layer.1.v_cache 0.00000077 0.00114289 - layer.2.k_cache 0.00115688 0.70083298 - layer.2.v_cache 0.00000135 0.00171508 - layer.3.k_cache 0.00129955 0.74164859 - layer.3.v_cache 0.00000214 0.00275408 - layer.4.k_cache 0.00368945 1.37543757 - layer.4.v_cache 0.00000337 0.00467829 - layer.4.output 0.00015649 0.09111335 - ------------------------------------------------------------------------------------- - TOTAL 0.00237842 1.63563688 - (elements=4,917,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4917248 -Total Bytes 599792 -BPFP 0.9758 bits/point -EBPFP 1.9516 equivalent bits/point -MSE 1.635637 ----------------------- -------------------------------------------------------- -Time: 1.092s Load: 0.018s, Pack+Encode: 0.399s, Decode+Unpack: 0.675s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 343, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 343, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6356 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample43-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample43-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample44-layer4-item1.zst (50/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample44-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 326, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 326, 128) -Output shape: (1, 326, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.0.v_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.1.k_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.1.v_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.2.k_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.2.v_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.3.k_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.3.v_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.4.k_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.4.v_cache: torch.Size([1, 8, 326, 128]) -> torch.Size([1, 1, 326, 1024]) - layer.4.output: torch.Size([1, 326, 4096]) -> torch.Size([1, 1, 326, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,788B, BPFP=0.2106 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,760B, BPFP=1.3602 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,496B, BPFP=0.6589 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 60,940B, BPFP=1.4604 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,720B, BPFP=0.8321 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 62,556B, BPFP=1.4991 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,956B, BPFP=0.9096 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 60,624B, BPFP=1.4528 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,776B, BPFP=0.6417 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 62,864B, BPFP=1.5065 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 143,452B, BPFP=0.8594 -⌛️ [2/4] FRONTEND: Frontend time: 0.405s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 326, 128]) - layer.0.v_cache: torch.Size([1, 8, 326, 128]) - layer.1.k_cache: torch.Size([1, 8, 326, 128]) - layer.1.v_cache: torch.Size([1, 8, 326, 128]) - layer.2.k_cache: torch.Size([1, 8, 326, 128]) - layer.2.v_cache: torch.Size([1, 8, 326, 128]) - layer.3.k_cache: torch.Size([1, 8, 326, 128]) - layer.3.v_cache: torch.Size([1, 8, 326, 128]) - layer.4.k_cache: torch.Size([1, 8, 326, 128]) - layer.4.v_cache: torch.Size([1, 8, 326, 128]) - layer.4.output: torch.Size([1, 326, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.672s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 326, 128]) - layer.0.v_cache: torch.Size([1, 8, 326, 128]) - layer.1.k_cache: torch.Size([1, 8, 326, 128]) - layer.1.v_cache: torch.Size([1, 8, 326, 128]) - layer.2.k_cache: torch.Size([1, 8, 326, 128]) - layer.2.v_cache: torch.Size([1, 8, 326, 128]) - layer.3.k_cache: torch.Size([1, 8, 326, 128]) - layer.3.v_cache: torch.Size([1, 8, 326, 128]) - layer.4.k_cache: torch.Size([1, 8, 326, 128]) - layer.4.v_cache: torch.Size([1, 8, 326, 128]) - layer.4.output: torch.Size([1, 326, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02438200 19.77468217 - layer.0.v_cache 0.00000027 0.00030697 - layer.1.k_cache 0.00290854 1.19321268 - layer.1.v_cache 0.00000077 0.00109210 - layer.2.k_cache 0.00114310 0.68434939 - layer.2.v_cache 0.00000112 0.00161401 - layer.3.k_cache 0.00131875 0.75095634 - layer.3.v_cache 0.00000210 0.00264104 - layer.4.k_cache 0.00360283 1.39369043 - layer.4.v_cache 0.00000318 0.00444391 - layer.4.output 0.00014710 0.08864836 - ------------------------------------------------------------------------------------- - TOTAL 0.00242508 1.72582732 - (elements=4,673,536) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4673536 -Total Bytes 582932 -BPFP 0.9978 bits/point -EBPFP 1.9957 equivalent bits/point -MSE 1.725827 ----------------------- -------------------------------------------------------- -Time: 1.094s Load: 0.016s, Pack+Encode: 0.405s, Decode+Unpack: 0.672s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 326, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 326, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7258 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample44-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample44-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample45-layer4-item1.zst (51/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample45-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 335, 128) -Output shape: (1, 335, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.output: torch.Size([1, 335, 4096]) -> torch.Size([1, 1, 335, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,324B, BPFP=0.2174 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,664B, BPFP=1.3681 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,116B, BPFP=0.6557 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,064B, BPFP=1.4707 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,096B, BPFP=0.8418 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,192B, BPFP=1.4970 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,880B, BPFP=0.9067 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,436B, BPFP=1.4561 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,676B, BPFP=0.6454 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,796B, BPFP=1.5111 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 151,920B, BPFP=0.8857 -⌛️ [2/4] FRONTEND: Frontend time: 0.401s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.680s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02422888 18.81008191 - layer.0.v_cache 0.00000026 0.00030539 - layer.1.k_cache 0.00291837 1.23278781 - layer.1.v_cache 0.00000081 0.00116220 - layer.2.k_cache 0.00116761 0.70162007 - layer.2.v_cache 0.00000118 0.00171098 - layer.3.k_cache 0.00133299 0.76708019 - layer.3.v_cache 0.00000217 0.00278798 - layer.4.k_cache 0.00352932 1.40029534 - layer.4.v_cache 0.00000340 0.00465047 - layer.4.output 0.00015509 0.09473073 - ------------------------------------------------------------------------------------- - TOTAL 0.00241467 1.66438609 - (elements=4,802,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4802560 -Total Bytes 605164 -BPFP 1.0081 bits/point -EBPFP 2.0161 equivalent bits/point -MSE 1.664386 ----------------------- -------------------------------------------------------- -Time: 1.099s Load: 0.018s, Pack+Encode: 0.401s, Decode+Unpack: 0.680s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6644 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample45-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample45-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample46-layer4-item1.zst (52/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample46-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 324, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 324, 128) -Output shape: (1, 324, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.0.v_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.1.k_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.1.v_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.2.k_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.2.v_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.3.k_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.3.v_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.4.k_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.4.v_cache: torch.Size([1, 8, 324, 128]) -> torch.Size([1, 1, 324, 1024]) - layer.4.output: torch.Size([1, 324, 4096]) -> torch.Size([1, 1, 324, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,728B, BPFP=0.2105 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 55,640B, BPFP=1.3416 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,440B, BPFP=0.6617 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 60,304B, BPFP=1.4541 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,416B, BPFP=0.8540 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 61,260B, BPFP=1.4771 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,372B, BPFP=0.9011 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 59,784B, BPFP=1.4416 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,552B, BPFP=0.6644 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 62,080B, BPFP=1.4969 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 147,416B, BPFP=0.8886 -⌛️ [2/4] FRONTEND: Frontend time: 0.398s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 324, 128]) - layer.0.v_cache: torch.Size([1, 8, 324, 128]) - layer.1.k_cache: torch.Size([1, 8, 324, 128]) - layer.1.v_cache: torch.Size([1, 8, 324, 128]) - layer.2.k_cache: torch.Size([1, 8, 324, 128]) - layer.2.v_cache: torch.Size([1, 8, 324, 128]) - layer.3.k_cache: torch.Size([1, 8, 324, 128]) - layer.3.v_cache: torch.Size([1, 8, 324, 128]) - layer.4.k_cache: torch.Size([1, 8, 324, 128]) - layer.4.v_cache: torch.Size([1, 8, 324, 128]) - layer.4.output: torch.Size([1, 324, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.677s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 324, 128]) - layer.0.v_cache: torch.Size([1, 8, 324, 128]) - layer.1.k_cache: torch.Size([1, 8, 324, 128]) - layer.1.v_cache: torch.Size([1, 8, 324, 128]) - layer.2.k_cache: torch.Size([1, 8, 324, 128]) - layer.2.v_cache: torch.Size([1, 8, 324, 128]) - layer.3.k_cache: torch.Size([1, 8, 324, 128]) - layer.3.v_cache: torch.Size([1, 8, 324, 128]) - layer.4.k_cache: torch.Size([1, 8, 324, 128]) - layer.4.v_cache: torch.Size([1, 8, 324, 128]) - layer.4.output: torch.Size([1, 324, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02313130 19.30700985 - layer.0.v_cache 0.00000026 0.00030165 - layer.1.k_cache 0.00295515 1.22524893 - layer.1.v_cache 0.00000079 0.00111747 - layer.2.k_cache 0.00117180 0.70636405 - layer.2.v_cache 0.00000115 0.00167953 - layer.3.k_cache 0.00132196 0.75926095 - layer.3.v_cache 0.00000215 0.00275900 - layer.4.k_cache 0.00355989 1.37740910 - layer.4.v_cache 0.00000355 0.00462576 - layer.4.output 0.00015612 0.09267766 - ------------------------------------------------------------------------------------- - TOTAL 0.00234089 1.69689193 - (elements=4,644,864) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4644864 -Total Bytes 582992 -BPFP 1.0041 bits/point -EBPFP 2.0082 equivalent bits/point -MSE 1.696892 ----------------------- -------------------------------------------------------- -Time: 1.093s Load: 0.017s, Pack+Encode: 0.398s, Decode+Unpack: 0.677s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 324, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 324, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6969 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample46-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample46-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample47-layer4-item1.zst (53/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample47-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 332, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 332, 128) -Output shape: (1, 332, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.0.v_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.1.k_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.1.v_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.2.k_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.2.v_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.3.k_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.3.v_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.4.k_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.4.v_cache: torch.Size([1, 8, 332, 128]) -> torch.Size([1, 1, 332, 1024]) - layer.4.output: torch.Size([1, 332, 4096]) -> torch.Size([1, 1, 332, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,016B, BPFP=0.2122 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,292B, BPFP=1.3482 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,216B, BPFP=0.6640 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,572B, BPFP=1.4724 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,012B, BPFP=0.8474 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,336B, BPFP=1.5139 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,428B, BPFP=0.9043 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,680B, BPFP=1.4750 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,012B, BPFP=0.6356 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,744B, BPFP=1.5235 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,244B, BPFP=0.8486 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 332, 128]) - layer.0.v_cache: torch.Size([1, 8, 332, 128]) - layer.1.k_cache: torch.Size([1, 8, 332, 128]) - layer.1.v_cache: torch.Size([1, 8, 332, 128]) - layer.2.k_cache: torch.Size([1, 8, 332, 128]) - layer.2.v_cache: torch.Size([1, 8, 332, 128]) - layer.3.k_cache: torch.Size([1, 8, 332, 128]) - layer.3.v_cache: torch.Size([1, 8, 332, 128]) - layer.4.k_cache: torch.Size([1, 8, 332, 128]) - layer.4.v_cache: torch.Size([1, 8, 332, 128]) - layer.4.output: torch.Size([1, 332, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.677s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 332, 128]) - layer.0.v_cache: torch.Size([1, 8, 332, 128]) - layer.1.k_cache: torch.Size([1, 8, 332, 128]) - layer.1.v_cache: torch.Size([1, 8, 332, 128]) - layer.2.k_cache: torch.Size([1, 8, 332, 128]) - layer.2.v_cache: torch.Size([1, 8, 332, 128]) - layer.3.k_cache: torch.Size([1, 8, 332, 128]) - layer.3.v_cache: torch.Size([1, 8, 332, 128]) - layer.4.k_cache: torch.Size([1, 8, 332, 128]) - layer.4.v_cache: torch.Size([1, 8, 332, 128]) - layer.4.output: torch.Size([1, 332, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02360621 19.63841597 - layer.0.v_cache 0.00000026 0.00030161 - layer.1.k_cache 0.00288168 1.18485269 - layer.1.v_cache 0.00000079 0.00115706 - layer.2.k_cache 0.00118772 0.71137766 - layer.2.v_cache 0.00000117 0.00175877 - layer.3.k_cache 0.00132341 0.76830738 - layer.3.v_cache 0.00000218 0.00283562 - layer.4.k_cache 0.00354949 1.42803256 - layer.4.v_cache 0.00000378 0.00485799 - layer.4.output 0.00018345 0.09609459 - ------------------------------------------------------------------------------------- - TOTAL 0.00237789 1.72330541 - (elements=4,759,552) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4759552 -Total Bytes 594552 -BPFP 0.9993 bits/point -EBPFP 1.9987 equivalent bits/point -MSE 1.723305 ----------------------- -------------------------------------------------------- -Time: 1.096s Load: 0.017s, Pack+Encode: 0.402s, Decode+Unpack: 0.677s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 332, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 332, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7233 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample47-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample47-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample48-layer4-item1.zst (54/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample48-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 319, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 319, 128) -Output shape: (1, 319, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.0.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.1.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.1.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.2.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.2.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.3.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.3.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.4.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.4.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.4.output: torch.Size([1, 319, 4096]) -> torch.Size([1, 1, 319, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,980B, BPFP=0.1954 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,076B, BPFP=1.2264 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,572B, BPFP=0.5773 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,956B, BPFP=1.3214 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,588B, BPFP=0.7736 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 54,996B, BPFP=1.3469 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,564B, BPFP=0.8220 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,324B, BPFP=1.3059 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,056B, BPFP=0.5647 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,364B, BPFP=1.3559 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 116,996B, BPFP=0.7163 -⌛️ [2/4] FRONTEND: Frontend time: 0.355s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 319, 128]) - layer.0.v_cache: torch.Size([1, 8, 319, 128]) - layer.1.k_cache: torch.Size([1, 8, 319, 128]) - layer.1.v_cache: torch.Size([1, 8, 319, 128]) - layer.2.k_cache: torch.Size([1, 8, 319, 128]) - layer.2.v_cache: torch.Size([1, 8, 319, 128]) - layer.3.k_cache: torch.Size([1, 8, 319, 128]) - layer.3.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.k_cache: torch.Size([1, 8, 319, 128]) - layer.4.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.output: torch.Size([1, 319, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.588s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 319, 128]) - layer.0.v_cache: torch.Size([1, 8, 319, 128]) - layer.1.k_cache: torch.Size([1, 8, 319, 128]) - layer.1.v_cache: torch.Size([1, 8, 319, 128]) - layer.2.k_cache: torch.Size([1, 8, 319, 128]) - layer.2.v_cache: torch.Size([1, 8, 319, 128]) - layer.3.k_cache: torch.Size([1, 8, 319, 128]) - layer.3.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.k_cache: torch.Size([1, 8, 319, 128]) - layer.4.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.output: torch.Size([1, 319, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02438342 17.34167442 - layer.0.v_cache 0.00000027 0.00030769 - layer.1.k_cache 0.00287626 1.11785544 - layer.1.v_cache 0.00000082 0.00115752 - layer.2.k_cache 0.00122126 0.67153452 - layer.2.v_cache 0.00000114 0.00167993 - layer.3.k_cache 0.00133159 0.75320851 - layer.3.v_cache 0.00000213 0.00270248 - layer.4.k_cache 0.00352266 1.36063206 - layer.4.v_cache 0.00000325 0.00461224 - layer.4.output 0.00013715 0.08726448 - ------------------------------------------------------------------------------------- - TOTAL 0.00242081 1.54317305 - (elements=4,573,184) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4573184 -Total Bytes 504472 -BPFP 0.8825 bits/point -EBPFP 1.7650 equivalent bits/point -MSE 1.543173 ----------------------- -------------------------------------------------------- -Time: 0.961s Load: 0.018s, Pack+Encode: 0.355s, Decode+Unpack: 0.588s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 319, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5432 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample48-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample48-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample49-layer4-item1.zst (55/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample49-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 335, 128) -Output shape: (1, 335, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) -> torch.Size([1, 1, 335, 1024]) - layer.4.output: torch.Size([1, 335, 4096]) -> torch.Size([1, 1, 335, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,168B, BPFP=0.2138 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,316B, BPFP=1.3600 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,908B, BPFP=0.6508 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,680B, BPFP=1.4618 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,440B, BPFP=0.8498 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,344B, BPFP=1.5006 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,764B, BPFP=0.9040 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,316B, BPFP=1.4533 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,800B, BPFP=0.6483 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,692B, BPFP=1.5087 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 143,348B, BPFP=0.8358 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.693s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 335, 128]) - layer.0.v_cache: torch.Size([1, 8, 335, 128]) - layer.1.k_cache: torch.Size([1, 8, 335, 128]) - layer.1.v_cache: torch.Size([1, 8, 335, 128]) - layer.2.k_cache: torch.Size([1, 8, 335, 128]) - layer.2.v_cache: torch.Size([1, 8, 335, 128]) - layer.3.k_cache: torch.Size([1, 8, 335, 128]) - layer.3.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.k_cache: torch.Size([1, 8, 335, 128]) - layer.4.v_cache: torch.Size([1, 8, 335, 128]) - layer.4.output: torch.Size([1, 335, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02322828 18.37181087 - layer.0.v_cache 0.00000026 0.00030360 - layer.1.k_cache 0.00283722 1.21996706 - layer.1.v_cache 0.00000078 0.00113773 - layer.2.k_cache 0.00117890 0.69881783 - layer.2.v_cache 0.00000113 0.00164131 - layer.3.k_cache 0.00131429 0.75652060 - layer.3.v_cache 0.00000212 0.00271302 - layer.4.k_cache 0.00357066 1.38745135 - layer.4.v_cache 0.00000341 0.00462969 - layer.4.output 0.00013949 0.09418015 - ------------------------------------------------------------------------------------- - TOTAL 0.00233536 1.63012241 - (elements=4,802,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4802560 -Total Bytes 595776 -BPFP 0.9924 bits/point -EBPFP 1.9849 equivalent bits/point -MSE 1.630122 ----------------------- -------------------------------------------------------- -Time: 1.117s Load: 0.017s, Pack+Encode: 0.407s, Decode+Unpack: 0.693s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 335, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 335, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6301 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample49-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample49-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample5-layer4-item1.zst (56/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample5-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 354, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 354, 128) -Output shape: (1, 354, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.0.v_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.1.k_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.1.v_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.2.k_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.2.v_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.3.k_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.3.v_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.4.k_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.4.v_cache: torch.Size([1, 8, 354, 128]) -> torch.Size([1, 1, 354, 1024]) - layer.4.output: torch.Size([1, 354, 4096]) -> torch.Size([1, 1, 354, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,320B, BPFP=0.2057 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,328B, BPFP=1.3093 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,944B, BPFP=0.6167 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,768B, BPFP=1.4073 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,232B, BPFP=0.7996 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,988B, BPFP=1.4122 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,536B, BPFP=0.8725 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,392B, BPFP=1.3769 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,340B, BPFP=0.6034 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,204B, BPFP=1.4390 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 147,128B, BPFP=0.8117 -⌛️ [2/4] FRONTEND: Frontend time: 0.412s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 354, 128]) - layer.0.v_cache: torch.Size([1, 8, 354, 128]) - layer.1.k_cache: torch.Size([1, 8, 354, 128]) - layer.1.v_cache: torch.Size([1, 8, 354, 128]) - layer.2.k_cache: torch.Size([1, 8, 354, 128]) - layer.2.v_cache: torch.Size([1, 8, 354, 128]) - layer.3.k_cache: torch.Size([1, 8, 354, 128]) - layer.3.v_cache: torch.Size([1, 8, 354, 128]) - layer.4.k_cache: torch.Size([1, 8, 354, 128]) - layer.4.v_cache: torch.Size([1, 8, 354, 128]) - layer.4.output: torch.Size([1, 354, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.692s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 354, 128]) - layer.0.v_cache: torch.Size([1, 8, 354, 128]) - layer.1.k_cache: torch.Size([1, 8, 354, 128]) - layer.1.v_cache: torch.Size([1, 8, 354, 128]) - layer.2.k_cache: torch.Size([1, 8, 354, 128]) - layer.2.v_cache: torch.Size([1, 8, 354, 128]) - layer.3.k_cache: torch.Size([1, 8, 354, 128]) - layer.3.v_cache: torch.Size([1, 8, 354, 128]) - layer.4.k_cache: torch.Size([1, 8, 354, 128]) - layer.4.v_cache: torch.Size([1, 8, 354, 128]) - layer.4.output: torch.Size([1, 354, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02475441 18.73963161 - layer.0.v_cache 0.00000027 0.00031808 - layer.1.k_cache 0.00288960 1.21944561 - layer.1.v_cache 0.00000085 0.00117142 - layer.2.k_cache 0.00116959 0.70241223 - layer.2.v_cache 0.00000115 0.00170804 - layer.3.k_cache 0.00132243 0.76917944 - layer.3.v_cache 0.00000218 0.00283169 - layer.4.k_cache 0.00355194 1.39653041 - layer.4.v_cache 0.00000329 0.00479681 - layer.4.output 0.00015235 0.09501465 - ------------------------------------------------------------------------------------- - TOTAL 0.00245037 1.65843457 - (elements=5,074,944) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5074944 -Total Bytes 602180 -BPFP 0.9493 bits/point -EBPFP 1.8985 equivalent bits/point -MSE 1.658435 ----------------------- -------------------------------------------------------- -Time: 1.123s Load: 0.019s, Pack+Encode: 0.412s, Decode+Unpack: 0.692s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 354, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 354, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6584 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample5-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample5-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample50-layer4-item1.zst (57/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample50-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 327, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 327, 128) -Output shape: (1, 327, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.0.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.1.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.1.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.2.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.2.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.3.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.3.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.4.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.4.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.4.output: torch.Size([1, 327, 4096]) -> torch.Size([1, 1, 327, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,892B, BPFP=0.2124 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,112B, BPFP=1.3645 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,032B, BPFP=0.6697 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,592B, BPFP=1.4715 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,604B, BPFP=0.8506 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 62,908B, BPFP=1.5030 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,372B, BPFP=0.9168 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,292B, BPFP=1.4644 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,808B, BPFP=0.6405 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,620B, BPFP=1.5200 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 149,632B, BPFP=0.8937 -⌛️ [2/4] FRONTEND: Frontend time: 0.414s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 327, 128]) - layer.0.v_cache: torch.Size([1, 8, 327, 128]) - layer.1.k_cache: torch.Size([1, 8, 327, 128]) - layer.1.v_cache: torch.Size([1, 8, 327, 128]) - layer.2.k_cache: torch.Size([1, 8, 327, 128]) - layer.2.v_cache: torch.Size([1, 8, 327, 128]) - layer.3.k_cache: torch.Size([1, 8, 327, 128]) - layer.3.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.k_cache: torch.Size([1, 8, 327, 128]) - layer.4.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.output: torch.Size([1, 327, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.694s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 327, 128]) - layer.0.v_cache: torch.Size([1, 8, 327, 128]) - layer.1.k_cache: torch.Size([1, 8, 327, 128]) - layer.1.v_cache: torch.Size([1, 8, 327, 128]) - layer.2.k_cache: torch.Size([1, 8, 327, 128]) - layer.2.v_cache: torch.Size([1, 8, 327, 128]) - layer.3.k_cache: torch.Size([1, 8, 327, 128]) - layer.3.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.k_cache: torch.Size([1, 8, 327, 128]) - layer.4.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.output: torch.Size([1, 327, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02504845 19.55337944 - layer.0.v_cache 0.00000027 0.00030227 - layer.1.k_cache 0.00290162 1.24400661 - layer.1.v_cache 0.00000083 0.00114636 - layer.2.k_cache 0.00115956 0.70170509 - layer.2.v_cache 0.00000120 0.00168247 - layer.3.k_cache 0.00131802 0.76351826 - layer.3.v_cache 0.00000223 0.00269624 - layer.4.k_cache 0.00351161 1.40452291 - layer.4.v_cache 0.00000355 0.00462090 - layer.4.output 0.00016630 0.10501981 - ------------------------------------------------------------------------------------- - TOTAL 0.00247233 1.72126141 - (elements=4,687,872) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4687872 -Total Bytes 593864 -BPFP 1.0134 bits/point -EBPFP 2.0269 equivalent bits/point -MSE 1.721261 ----------------------- -------------------------------------------------------- -Time: 1.124s Load: 0.017s, Pack+Encode: 0.414s, Decode+Unpack: 0.694s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 327, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7213 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample50-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample50-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample51-layer4-item1.zst (58/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample51-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 313, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.015s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 313, 128) -Output shape: (1, 313, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.0.v_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.1.k_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.1.v_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.2.k_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.2.v_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.3.k_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.3.v_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.4.k_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.4.v_cache: torch.Size([1, 8, 313, 128]) -> torch.Size([1, 1, 313, 1024]) - layer.4.output: torch.Size([1, 313, 4096]) -> torch.Size([1, 1, 313, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,112B, BPFP=0.2025 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,848B, BPFP=1.2692 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,984B, BPFP=0.5986 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 55,076B, BPFP=1.3747 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,900B, BPFP=0.7962 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,516B, BPFP=1.3857 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,044B, BPFP=0.8497 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 54,040B, BPFP=1.3488 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,536B, BPFP=0.5875 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,232B, BPFP=1.4036 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 134,068B, BPFP=0.8366 -⌛️ [2/4] FRONTEND: Frontend time: 0.353s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 313, 128]) - layer.0.v_cache: torch.Size([1, 8, 313, 128]) - layer.1.k_cache: torch.Size([1, 8, 313, 128]) - layer.1.v_cache: torch.Size([1, 8, 313, 128]) - layer.2.k_cache: torch.Size([1, 8, 313, 128]) - layer.2.v_cache: torch.Size([1, 8, 313, 128]) - layer.3.k_cache: torch.Size([1, 8, 313, 128]) - layer.3.v_cache: torch.Size([1, 8, 313, 128]) - layer.4.k_cache: torch.Size([1, 8, 313, 128]) - layer.4.v_cache: torch.Size([1, 8, 313, 128]) - layer.4.output: torch.Size([1, 313, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.590s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 313, 128]) - layer.0.v_cache: torch.Size([1, 8, 313, 128]) - layer.1.k_cache: torch.Size([1, 8, 313, 128]) - layer.1.v_cache: torch.Size([1, 8, 313, 128]) - layer.2.k_cache: torch.Size([1, 8, 313, 128]) - layer.2.v_cache: torch.Size([1, 8, 313, 128]) - layer.3.k_cache: torch.Size([1, 8, 313, 128]) - layer.3.v_cache: torch.Size([1, 8, 313, 128]) - layer.4.k_cache: torch.Size([1, 8, 313, 128]) - layer.4.v_cache: torch.Size([1, 8, 313, 128]) - layer.4.output: torch.Size([1, 313, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02366765 19.29529784 - layer.0.v_cache 0.00000027 0.00029912 - layer.1.k_cache 0.00295685 1.15667832 - layer.1.v_cache 0.00000081 0.00106205 - layer.2.k_cache 0.00117427 0.69147528 - layer.2.v_cache 0.00000111 0.00161689 - layer.3.k_cache 0.00131811 0.75296698 - layer.3.v_cache 0.00000214 0.00264104 - layer.4.k_cache 0.00360037 1.38988926 - layer.4.v_cache 0.00000322 0.00453196 - layer.4.output 0.00015721 0.10090166 - ------------------------------------------------------------------------------------- - TOTAL 0.00238240 1.69286181 - (elements=4,487,168) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4487168 -Total Bytes 527356 -BPFP 0.9402 bits/point -EBPFP 1.8804 equivalent bits/point -MSE 1.692862 ----------------------- -------------------------------------------------------- -Time: 0.958s Load: 0.015s, Pack+Encode: 0.353s, Decode+Unpack: 0.590s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 313, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 313, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6929 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample51-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample51-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample52-layer4-item1.zst (59/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample52-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 329, 128) -Output shape: (1, 329, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.output: torch.Size([1, 329, 4096]) -> torch.Size([1, 1, 329, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,876B, BPFP=0.2108 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,964B, BPFP=1.3527 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,628B, BPFP=0.6561 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,876B, BPFP=1.4693 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,484B, BPFP=0.8426 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,612B, BPFP=1.5105 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,116B, BPFP=0.9051 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,796B, BPFP=1.4674 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,572B, BPFP=0.6310 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,996B, BPFP=1.5197 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 152,068B, BPFP=0.9028 -⌛️ [2/4] FRONTEND: Frontend time: 0.404s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.684s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02435147 19.83768878 - layer.0.v_cache 0.00000027 0.00030976 - layer.1.k_cache 0.00282437 1.20142529 - layer.1.v_cache 0.00000081 0.00117328 - layer.2.k_cache 0.00117800 0.70216045 - layer.2.v_cache 0.00000117 0.00172598 - layer.3.k_cache 0.00131145 0.77617879 - layer.3.v_cache 0.00000228 0.00284953 - layer.4.k_cache 0.00354228 1.43889305 - layer.4.v_cache 0.00000348 0.00481627 - layer.4.output 0.00016143 0.09867059 - ------------------------------------------------------------------------------------- - TOTAL 0.00241866 1.74013597 - (elements=4,716,544) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4716544 -Total Bytes 596988 -BPFP 1.0126 bits/point -EBPFP 2.0252 equivalent bits/point -MSE 1.740136 ----------------------- -------------------------------------------------------- -Time: 1.105s Load: 0.017s, Pack+Encode: 0.404s, Decode+Unpack: 0.684s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7401 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample52-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample52-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample53-layer4-item1.zst (60/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample53-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 330, 128) -Output shape: (1, 330, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.output: torch.Size([1, 330, 4096]) -> torch.Size([1, 1, 330, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,024B, BPFP=0.2136 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,608B, BPFP=1.3638 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,132B, BPFP=0.6660 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,676B, BPFP=1.4838 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,088B, BPFP=0.8544 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,260B, BPFP=1.5213 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,564B, BPFP=0.9130 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,512B, BPFP=1.4799 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,304B, BPFP=0.6701 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,828B, BPFP=1.5348 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,164B, BPFP=0.8769 -⌛️ [2/4] FRONTEND: Frontend time: 0.405s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.683s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02327405 20.47722686 - layer.0.v_cache 0.00000026 0.00030660 - layer.1.k_cache 0.00286502 1.20881616 - layer.1.v_cache 0.00000078 0.00115004 - layer.2.k_cache 0.00115867 0.70328133 - layer.2.v_cache 0.00000116 0.00169689 - layer.3.k_cache 0.00132713 0.76823980 - layer.3.v_cache 0.00000212 0.00273596 - layer.4.k_cache 0.00352212 1.40831733 - layer.4.v_cache 0.00000405 0.00480107 - layer.4.output 0.00014449 0.09638840 - ------------------------------------------------------------------------------------- - TOTAL 0.00233809 1.78300898 - (elements=4,730,880) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4730880 -Total Bytes 600160 -BPFP 1.0149 bits/point -EBPFP 2.0298 equivalent bits/point -MSE 1.783009 ----------------------- -------------------------------------------------------- -Time: 1.105s Load: 0.017s, Pack+Encode: 0.405s, Decode+Unpack: 0.683s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7830 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample53-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample53-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample54-layer4-item1.zst (61/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample54-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 320, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 320, 128) -Output shape: (1, 320, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.0.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.1.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.1.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.2.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.2.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.3.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.3.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.4.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.4.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.4.output: torch.Size([1, 320, 4096]) -> torch.Size([1, 1, 320, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,708B, BPFP=0.1882 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 49,356B, BPFP=1.2050 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,664B, BPFP=0.5533 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 52,980B, BPFP=1.2935 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,820B, BPFP=0.7524 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 54,396B, BPFP=1.3280 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,712B, BPFP=0.7986 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,012B, BPFP=1.2942 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,332B, BPFP=0.5452 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,284B, BPFP=1.3497 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 118,616B, BPFP=0.7240 -⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 320, 128]) - layer.0.v_cache: torch.Size([1, 8, 320, 128]) - layer.1.k_cache: torch.Size([1, 8, 320, 128]) - layer.1.v_cache: torch.Size([1, 8, 320, 128]) - layer.2.k_cache: torch.Size([1, 8, 320, 128]) - layer.2.v_cache: torch.Size([1, 8, 320, 128]) - layer.3.k_cache: torch.Size([1, 8, 320, 128]) - layer.3.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.k_cache: torch.Size([1, 8, 320, 128]) - layer.4.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.output: torch.Size([1, 320, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.584s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 320, 128]) - layer.0.v_cache: torch.Size([1, 8, 320, 128]) - layer.1.k_cache: torch.Size([1, 8, 320, 128]) - layer.1.v_cache: torch.Size([1, 8, 320, 128]) - layer.2.k_cache: torch.Size([1, 8, 320, 128]) - layer.2.v_cache: torch.Size([1, 8, 320, 128]) - layer.3.k_cache: torch.Size([1, 8, 320, 128]) - layer.3.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.k_cache: torch.Size([1, 8, 320, 128]) - layer.4.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.output: torch.Size([1, 320, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02461678 17.24787598 - layer.0.v_cache 0.00000026 0.00029798 - layer.1.k_cache 0.00291374 1.06878757 - layer.1.v_cache 0.00000082 0.00111337 - layer.2.k_cache 0.00121384 0.68502016 - layer.2.v_cache 0.00000117 0.00169171 - layer.3.k_cache 0.00132315 0.72866564 - layer.3.v_cache 0.00000219 0.00276333 - layer.4.k_cache 0.00353222 1.31388874 - layer.4.v_cache 0.00000326 0.00471275 - layer.4.output 0.00015660 0.08598354 - ------------------------------------------------------------------------------------- - TOTAL 0.00244527 1.52848224 - (elements=4,587,520) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4587520 -Total Bytes 499880 -BPFP 0.8717 bits/point -EBPFP 1.7434 equivalent bits/point -MSE 1.528482 ----------------------- -------------------------------------------------------- -Time: 0.956s Load: 0.018s, Pack+Encode: 0.354s, Decode+Unpack: 0.584s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 320, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5285 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample54-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample54-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample55-layer4-item1.zst (62/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample55-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 339, 128) -Output shape: (1, 339, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.output: torch.Size([1, 339, 4096]) -> torch.Size([1, 1, 339, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,336B, BPFP=0.2152 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,504B, BPFP=1.3483 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,212B, BPFP=0.6502 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,496B, BPFP=1.4633 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,652B, BPFP=0.8447 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,840B, BPFP=1.4943 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,172B, BPFP=0.9027 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,028B, BPFP=1.4525 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,476B, BPFP=0.6332 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,120B, BPFP=1.5007 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,444B, BPFP=0.8380 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.681s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02391626 18.40768604 - layer.0.v_cache 0.00000026 0.00030485 - layer.1.k_cache 0.00286983 1.23642697 - layer.1.v_cache 0.00000081 0.00116148 - layer.2.k_cache 0.00115809 0.70388546 - layer.2.v_cache 0.00000117 0.00172352 - layer.3.k_cache 0.00131372 0.77026961 - layer.3.v_cache 0.00000240 0.00282068 - layer.4.k_cache 0.00359509 1.41823657 - layer.4.v_cache 0.00000368 0.00479933 - layer.4.output 0.00014345 0.09241976 - ------------------------------------------------------------------------------------- - TOTAL 0.00238822 1.63692811 - (elements=4,859,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4859904 -Total Bytes 601280 -BPFP 0.9898 bits/point -EBPFP 1.9796 equivalent bits/point -MSE 1.636928 ----------------------- -------------------------------------------------------- -Time: 1.107s Load: 0.019s, Pack+Encode: 0.407s, Decode+Unpack: 0.681s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6369 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample55-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample55-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample56-layer4-item1.zst (63/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample56-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 314, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 314, 128) -Output shape: (1, 314, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.0.v_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.1.k_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.1.v_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.2.k_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.2.v_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.3.k_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.3.v_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.4.k_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.4.v_cache: torch.Size([1, 8, 314, 128]) -> torch.Size([1, 1, 314, 1024]) - layer.4.output: torch.Size([1, 314, 4096]) -> torch.Size([1, 1, 314, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,944B, BPFP=0.1977 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,752B, BPFP=1.2627 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,840B, BPFP=0.5932 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 55,216B, BPFP=1.3738 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,612B, BPFP=0.7865 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,348B, BPFP=1.3771 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,832B, BPFP=0.8418 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,912B, BPFP=1.3414 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,512B, BPFP=0.5850 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,220B, BPFP=1.3988 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 125,816B, BPFP=0.7826 -⌛️ [2/4] FRONTEND: Frontend time: 0.374s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 314, 128]) - layer.0.v_cache: torch.Size([1, 8, 314, 128]) - layer.1.k_cache: torch.Size([1, 8, 314, 128]) - layer.1.v_cache: torch.Size([1, 8, 314, 128]) - layer.2.k_cache: torch.Size([1, 8, 314, 128]) - layer.2.v_cache: torch.Size([1, 8, 314, 128]) - layer.3.k_cache: torch.Size([1, 8, 314, 128]) - layer.3.v_cache: torch.Size([1, 8, 314, 128]) - layer.4.k_cache: torch.Size([1, 8, 314, 128]) - layer.4.v_cache: torch.Size([1, 8, 314, 128]) - layer.4.output: torch.Size([1, 314, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.588s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 314, 128]) - layer.0.v_cache: torch.Size([1, 8, 314, 128]) - layer.1.k_cache: torch.Size([1, 8, 314, 128]) - layer.1.v_cache: torch.Size([1, 8, 314, 128]) - layer.2.k_cache: torch.Size([1, 8, 314, 128]) - layer.2.v_cache: torch.Size([1, 8, 314, 128]) - layer.3.k_cache: torch.Size([1, 8, 314, 128]) - layer.3.v_cache: torch.Size([1, 8, 314, 128]) - layer.4.k_cache: torch.Size([1, 8, 314, 128]) - layer.4.v_cache: torch.Size([1, 8, 314, 128]) - layer.4.output: torch.Size([1, 314, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02380536 19.37961846 - layer.0.v_cache 0.00000027 0.00031072 - layer.1.k_cache 0.00292133 1.22447448 - layer.1.v_cache 0.00000080 0.00113719 - layer.2.k_cache 0.00120652 0.69281851 - layer.2.v_cache 0.00000113 0.00166840 - layer.3.k_cache 0.00133526 0.75407478 - layer.3.v_cache 0.00000215 0.00271356 - layer.4.k_cache 0.00353655 1.41681171 - layer.4.v_cache 0.00000336 0.00464252 - layer.4.output 0.00016938 0.09607892 - ------------------------------------------------------------------------------------- - TOTAL 0.00239216 1.70447043 - (elements=4,501,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4501504 -Total Bytes 518004 -BPFP 0.9206 bits/point -EBPFP 1.8412 equivalent bits/point -MSE 1.704470 ----------------------- -------------------------------------------------------- -Time: 0.978s Load: 0.016s, Pack+Encode: 0.374s, Decode+Unpack: 0.588s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 314, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 314, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7045 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample56-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample56-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample57-layer4-item1.zst (64/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample57-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 344, 128) -Output shape: (1, 344, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.output: torch.Size([1, 344, 4096]) -> torch.Size([1, 1, 344, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,216B, BPFP=0.2093 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,680B, BPFP=1.3554 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,100B, BPFP=0.6382 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,076B, BPFP=1.4552 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,932B, BPFP=0.8388 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,532B, BPFP=1.4883 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,444B, BPFP=0.8958 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,544B, BPFP=1.4431 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,944B, BPFP=0.6346 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,516B, BPFP=1.5106 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 142,960B, BPFP=0.8117 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.679s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02423059 19.25831781 - layer.0.v_cache 0.00000027 0.00030353 - layer.1.k_cache 0.00284490 1.20890533 - layer.1.v_cache 0.00000080 0.00116987 - layer.2.k_cache 0.00115895 0.68842054 - layer.2.v_cache 0.00000119 0.00174269 - layer.3.k_cache 0.00130300 0.74405435 - layer.3.v_cache 0.00000215 0.00283382 - layer.4.k_cache 0.00352154 1.35315341 - layer.4.v_cache 0.00000376 0.00482354 - layer.4.output 0.00013131 0.08325161 - ------------------------------------------------------------------------------------- - TOTAL 0.00239946 1.68548081 - (elements=4,931,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4931584 -Total Bytes 603944 -BPFP 0.9797 bits/point -EBPFP 1.9594 equivalent bits/point -MSE 1.685481 ----------------------- -------------------------------------------------------- -Time: 1.104s Load: 0.019s, Pack+Encode: 0.406s, Decode+Unpack: 0.679s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6855 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample57-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample57-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample58-layer4-item1.zst (65/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample58-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 308, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 308, 128) -Output shape: (1, 308, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.0.v_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.1.k_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.1.v_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.2.k_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.2.v_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.3.k_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.3.v_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.4.k_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.4.v_cache: torch.Size([1, 8, 308, 128]) -> torch.Size([1, 1, 308, 1024]) - layer.4.output: torch.Size([1, 308, 4096]) -> torch.Size([1, 1, 308, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,912B, BPFP=0.2007 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,108B, BPFP=1.2710 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,940B, BPFP=0.6072 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 54,316B, BPFP=1.3777 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,096B, BPFP=0.7888 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,212B, BPFP=1.4005 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,552B, BPFP=0.8511 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,484B, BPFP=1.3566 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,820B, BPFP=0.6042 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,860B, BPFP=1.4169 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 131,208B, BPFP=0.8320 -⌛️ [2/4] FRONTEND: Frontend time: 0.355s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 308, 128]) - layer.0.v_cache: torch.Size([1, 8, 308, 128]) - layer.1.k_cache: torch.Size([1, 8, 308, 128]) - layer.1.v_cache: torch.Size([1, 8, 308, 128]) - layer.2.k_cache: torch.Size([1, 8, 308, 128]) - layer.2.v_cache: torch.Size([1, 8, 308, 128]) - layer.3.k_cache: torch.Size([1, 8, 308, 128]) - layer.3.v_cache: torch.Size([1, 8, 308, 128]) - layer.4.k_cache: torch.Size([1, 8, 308, 128]) - layer.4.v_cache: torch.Size([1, 8, 308, 128]) - layer.4.output: torch.Size([1, 308, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.576s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 308, 128]) - layer.0.v_cache: torch.Size([1, 8, 308, 128]) - layer.1.k_cache: torch.Size([1, 8, 308, 128]) - layer.1.v_cache: torch.Size([1, 8, 308, 128]) - layer.2.k_cache: torch.Size([1, 8, 308, 128]) - layer.2.v_cache: torch.Size([1, 8, 308, 128]) - layer.3.k_cache: torch.Size([1, 8, 308, 128]) - layer.3.v_cache: torch.Size([1, 8, 308, 128]) - layer.4.k_cache: torch.Size([1, 8, 308, 128]) - layer.4.v_cache: torch.Size([1, 8, 308, 128]) - layer.4.output: torch.Size([1, 308, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02345257 18.71137822 - layer.0.v_cache 0.00000027 0.00030373 - layer.1.k_cache 0.00291045 1.20901578 - layer.1.v_cache 0.00000082 0.00115072 - layer.2.k_cache 0.00116968 0.69695703 - layer.2.v_cache 0.00000115 0.00167708 - layer.3.k_cache 0.00131855 0.76161778 - layer.3.v_cache 0.00000224 0.00280763 - layer.4.k_cache 0.00354502 1.41128114 - layer.4.v_cache 0.00000324 0.00471181 - layer.4.output 0.00015550 0.09529671 - ------------------------------------------------------------------------------------- - TOTAL 0.00235900 1.65586341 - (elements=4,415,488) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4415488 -Total Bytes 520508 -BPFP 0.9431 bits/point -EBPFP 1.8861 equivalent bits/point -MSE 1.655863 ----------------------- -------------------------------------------------------- -Time: 0.948s Load: 0.017s, Pack+Encode: 0.355s, Decode+Unpack: 0.576s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 308, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 308, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6559 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample58-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample58-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample59-layer4-item1.zst (66/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample59-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 339, 128) -Output shape: (1, 339, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) -> torch.Size([1, 1, 339, 1024]) - layer.4.output: torch.Size([1, 339, 4096]) -> torch.Size([1, 1, 339, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,056B, BPFP=0.2087 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,288B, BPFP=1.3433 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,092B, BPFP=0.6474 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,356B, BPFP=1.4601 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,384B, BPFP=0.8154 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,376B, BPFP=1.4836 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,664B, BPFP=0.8910 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,864B, BPFP=1.4487 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,752B, BPFP=0.6165 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,892B, BPFP=1.4955 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 146,712B, BPFP=0.8453 -⌛️ [2/4] FRONTEND: Frontend time: 0.411s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.694s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 339, 128]) - layer.0.v_cache: torch.Size([1, 8, 339, 128]) - layer.1.k_cache: torch.Size([1, 8, 339, 128]) - layer.1.v_cache: torch.Size([1, 8, 339, 128]) - layer.2.k_cache: torch.Size([1, 8, 339, 128]) - layer.2.v_cache: torch.Size([1, 8, 339, 128]) - layer.3.k_cache: torch.Size([1, 8, 339, 128]) - layer.3.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.k_cache: torch.Size([1, 8, 339, 128]) - layer.4.v_cache: torch.Size([1, 8, 339, 128]) - layer.4.output: torch.Size([1, 339, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02515205 19.65563353 - layer.0.v_cache 0.00000026 0.00030769 - layer.1.k_cache 0.00294413 1.23401905 - layer.1.v_cache 0.00000079 0.00114691 - layer.2.k_cache 0.00117757 0.70484992 - layer.2.v_cache 0.00000116 0.00171178 - layer.3.k_cache 0.00130963 0.77316887 - layer.3.v_cache 0.00000218 0.00284831 - layer.4.k_cache 0.00367572 1.41452936 - layer.4.v_cache 0.00000325 0.00469648 - layer.4.output 0.00014293 0.09088288 - ------------------------------------------------------------------------------------- - TOTAL 0.00248846 1.72546025 - (elements=4,859,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4859904 -Total Bytes 598436 -BPFP 0.9851 bits/point -EBPFP 1.9702 equivalent bits/point -MSE 1.725460 ----------------------- -------------------------------------------------------- -Time: 1.124s Load: 0.019s, Pack+Encode: 0.411s, Decode+Unpack: 0.694s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 339, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 339, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7255 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample59-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample59-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample6-layer4-item1.zst (67/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample6-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 355, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 355, 128) -Output shape: (1, 355, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.0.v_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.1.k_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.1.v_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.2.k_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.2.v_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.3.k_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.3.v_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.4.k_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.4.v_cache: torch.Size([1, 8, 355, 128]) -> torch.Size([1, 1, 355, 1024]) - layer.4.output: torch.Size([1, 355, 4096]) -> torch.Size([1, 1, 355, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,904B, BPFP=0.1960 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,740B, BPFP=1.3147 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,120B, BPFP=0.6188 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,084B, BPFP=1.4103 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,540B, BPFP=0.8261 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,636B, BPFP=1.4224 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,844B, BPFP=0.8768 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,528B, BPFP=1.3761 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,716B, BPFP=0.6099 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,556B, BPFP=1.4427 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 139,788B, BPFP=0.7691 -⌛️ [2/4] FRONTEND: Frontend time: 0.410s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 355, 128]) - layer.0.v_cache: torch.Size([1, 8, 355, 128]) - layer.1.k_cache: torch.Size([1, 8, 355, 128]) - layer.1.v_cache: torch.Size([1, 8, 355, 128]) - layer.2.k_cache: torch.Size([1, 8, 355, 128]) - layer.2.v_cache: torch.Size([1, 8, 355, 128]) - layer.3.k_cache: torch.Size([1, 8, 355, 128]) - layer.3.v_cache: torch.Size([1, 8, 355, 128]) - layer.4.k_cache: torch.Size([1, 8, 355, 128]) - layer.4.v_cache: torch.Size([1, 8, 355, 128]) - layer.4.output: torch.Size([1, 355, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.695s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 355, 128]) - layer.0.v_cache: torch.Size([1, 8, 355, 128]) - layer.1.k_cache: torch.Size([1, 8, 355, 128]) - layer.1.v_cache: torch.Size([1, 8, 355, 128]) - layer.2.k_cache: torch.Size([1, 8, 355, 128]) - layer.2.v_cache: torch.Size([1, 8, 355, 128]) - layer.3.k_cache: torch.Size([1, 8, 355, 128]) - layer.3.v_cache: torch.Size([1, 8, 355, 128]) - layer.4.k_cache: torch.Size([1, 8, 355, 128]) - layer.4.v_cache: torch.Size([1, 8, 355, 128]) - layer.4.output: torch.Size([1, 355, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02438693 18.13124175 - layer.0.v_cache 0.00000027 0.00030981 - layer.1.k_cache 0.00283684 1.19688256 - layer.1.v_cache 0.00000080 0.00119812 - layer.2.k_cache 0.00116829 0.68808327 - layer.2.v_cache 0.00000120 0.00176523 - layer.3.k_cache 0.00130356 0.75793302 - layer.3.v_cache 0.00000219 0.00281734 - layer.4.k_cache 0.00355311 1.39009537 - layer.4.v_cache 0.00000352 0.00476494 - layer.4.output 0.00013754 0.08616407 - ------------------------------------------------------------------------------------- - TOTAL 0.00241478 1.60855341 - (elements=5,089,280) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5089280 -Total Bytes 598456 -BPFP 0.9407 bits/point -EBPFP 1.8815 equivalent bits/point -MSE 1.608553 ----------------------- -------------------------------------------------------- -Time: 1.124s Load: 0.019s, Pack+Encode: 0.410s, Decode+Unpack: 0.695s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 355, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 355, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6086 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample6-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample6-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample60-layer4-item1.zst (68/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample60-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 334, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 334, 128) -Output shape: (1, 334, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.0.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.1.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.1.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.2.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.2.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.3.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.3.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.4.k_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.4.v_cache: torch.Size([1, 8, 334, 128]) -> torch.Size([1, 1, 334, 1024]) - layer.4.output: torch.Size([1, 334, 4096]) -> torch.Size([1, 1, 334, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,172B, BPFP=0.2145 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,904B, BPFP=1.3544 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,788B, BPFP=0.6500 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,496B, BPFP=1.4618 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,252B, BPFP=0.8480 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,924B, BPFP=1.4952 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,444B, BPFP=0.8992 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,584B, BPFP=1.4639 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,536B, BPFP=0.6441 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,820B, BPFP=1.5162 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 151,104B, BPFP=0.8836 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 334, 128]) - layer.0.v_cache: torch.Size([1, 8, 334, 128]) - layer.1.k_cache: torch.Size([1, 8, 334, 128]) - layer.1.v_cache: torch.Size([1, 8, 334, 128]) - layer.2.k_cache: torch.Size([1, 8, 334, 128]) - layer.2.v_cache: torch.Size([1, 8, 334, 128]) - layer.3.k_cache: torch.Size([1, 8, 334, 128]) - layer.3.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.k_cache: torch.Size([1, 8, 334, 128]) - layer.4.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.output: torch.Size([1, 334, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.690s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 334, 128]) - layer.0.v_cache: torch.Size([1, 8, 334, 128]) - layer.1.k_cache: torch.Size([1, 8, 334, 128]) - layer.1.v_cache: torch.Size([1, 8, 334, 128]) - layer.2.k_cache: torch.Size([1, 8, 334, 128]) - layer.2.v_cache: torch.Size([1, 8, 334, 128]) - layer.3.k_cache: torch.Size([1, 8, 334, 128]) - layer.3.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.k_cache: torch.Size([1, 8, 334, 128]) - layer.4.v_cache: torch.Size([1, 8, 334, 128]) - layer.4.output: torch.Size([1, 334, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02444427 19.29326113 - layer.0.v_cache 0.00000026 0.00031003 - layer.1.k_cache 0.00284576 1.21597107 - layer.1.v_cache 0.00000078 0.00117278 - layer.2.k_cache 0.00117801 0.70266102 - layer.2.v_cache 0.00000115 0.00168582 - layer.3.k_cache 0.00132187 0.76801122 - layer.3.v_cache 0.00000218 0.00284167 - layer.4.k_cache 0.00348374 1.40169228 - layer.4.v_cache 0.00000337 0.00485977 - layer.4.output 0.00014793 0.09168858 - ------------------------------------------------------------------------------------- - TOTAL 0.00241951 1.69708722 - (elements=4,788,224) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4788224 -Total Bytes 602024 -BPFP 1.0058 bits/point -EBPFP 2.0117 equivalent bits/point -MSE 1.697087 ----------------------- -------------------------------------------------------- -Time: 1.114s Load: 0.017s, Pack+Encode: 0.408s, Decode+Unpack: 0.690s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 334, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 334, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6971 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample60-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample60-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample61-layer4-item1.zst (69/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample61-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 342, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 342, 128) -Output shape: (1, 342, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.output: torch.Size([1, 342, 4096]) -> torch.Size([1, 1, 342, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,080B, BPFP=0.2074 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,108B, BPFP=1.3502 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,672B, BPFP=0.6321 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,228B, BPFP=1.4672 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,404B, BPFP=0.8316 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,016B, BPFP=1.4852 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,244B, BPFP=0.8965 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,160B, BPFP=1.4428 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,104B, BPFP=0.6192 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,584B, BPFP=1.4982 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 138,440B, BPFP=0.7906 -⌛️ [2/4] FRONTEND: Frontend time: 0.409s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.output: torch.Size([1, 342, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.682s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.output: torch.Size([1, 342, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02429651 18.83564339 - layer.0.v_cache 0.00000027 0.00030816 - layer.1.k_cache 0.00293320 1.20877602 - layer.1.v_cache 0.00000079 0.00118047 - layer.2.k_cache 0.00117175 0.71735311 - layer.2.v_cache 0.00000114 0.00172005 - layer.3.k_cache 0.00133100 0.76688300 - layer.3.v_cache 0.00000223 0.00272924 - layer.4.k_cache 0.00354789 1.41530944 - layer.4.v_cache 0.00000344 0.00461987 - layer.4.output 0.00015145 0.09137015 - ------------------------------------------------------------------------------------- - TOTAL 0.00242100 1.66571453 - (elements=4,902,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4902912 -Total Bytes 595040 -BPFP 0.9709 bits/point -EBPFP 1.9418 equivalent bits/point -MSE 1.665715 ----------------------- -------------------------------------------------------- -Time: 1.110s Load: 0.018s, Pack+Encode: 0.409s, Decode+Unpack: 0.682s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 342, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6657 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample61-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample61-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample62-layer4-item1.zst (70/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample62-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 312, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.015s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 312, 128) -Output shape: (1, 312, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.0.v_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.1.k_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.1.v_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.2.k_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.2.v_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.3.k_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.3.v_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.4.k_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.4.v_cache: torch.Size([1, 8, 312, 128]) -> torch.Size([1, 1, 312, 1024]) - layer.4.output: torch.Size([1, 312, 4096]) -> torch.Size([1, 1, 312, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,960B, BPFP=0.1993 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,052B, BPFP=1.2533 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,664B, BPFP=0.5925 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 54,468B, BPFP=1.3639 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,416B, BPFP=0.7867 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,188B, BPFP=1.3819 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,576B, BPFP=0.8407 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,660B, BPFP=1.3436 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,864B, BPFP=0.5725 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,540B, BPFP=1.3907 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 131,760B, BPFP=0.8248 -⌛️ [2/4] FRONTEND: Frontend time: 0.353s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 312, 128]) - layer.0.v_cache: torch.Size([1, 8, 312, 128]) - layer.1.k_cache: torch.Size([1, 8, 312, 128]) - layer.1.v_cache: torch.Size([1, 8, 312, 128]) - layer.2.k_cache: torch.Size([1, 8, 312, 128]) - layer.2.v_cache: torch.Size([1, 8, 312, 128]) - layer.3.k_cache: torch.Size([1, 8, 312, 128]) - layer.3.v_cache: torch.Size([1, 8, 312, 128]) - layer.4.k_cache: torch.Size([1, 8, 312, 128]) - layer.4.v_cache: torch.Size([1, 8, 312, 128]) - layer.4.output: torch.Size([1, 312, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.584s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 312, 128]) - layer.0.v_cache: torch.Size([1, 8, 312, 128]) - layer.1.k_cache: torch.Size([1, 8, 312, 128]) - layer.1.v_cache: torch.Size([1, 8, 312, 128]) - layer.2.k_cache: torch.Size([1, 8, 312, 128]) - layer.2.v_cache: torch.Size([1, 8, 312, 128]) - layer.3.k_cache: torch.Size([1, 8, 312, 128]) - layer.3.v_cache: torch.Size([1, 8, 312, 128]) - layer.4.k_cache: torch.Size([1, 8, 312, 128]) - layer.4.v_cache: torch.Size([1, 8, 312, 128]) - layer.4.output: torch.Size([1, 312, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02396220 19.47639661 - layer.0.v_cache 0.00000026 0.00030637 - layer.1.k_cache 0.00299662 1.17226928 - layer.1.v_cache 0.00000081 0.00111991 - layer.2.k_cache 0.00115233 0.70895009 - layer.2.v_cache 0.00000114 0.00171825 - layer.3.k_cache 0.00133293 0.75898689 - layer.3.v_cache 0.00000220 0.00280127 - layer.4.k_cache 0.00362032 1.41083654 - layer.4.v_cache 0.00000325 0.00456477 - layer.4.output 0.00015781 0.09618675 - ------------------------------------------------------------------------------------- - TOTAL 0.00240738 1.70876407 - (elements=4,472,832) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4472832 -Total Bytes 520148 -BPFP 0.9303 bits/point -EBPFP 1.8606 equivalent bits/point -MSE 1.708764 ----------------------- -------------------------------------------------------- -Time: 0.952s Load: 0.015s, Pack+Encode: 0.353s, Decode+Unpack: 0.584s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 312, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 312, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7088 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample62-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample62-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample63-layer4-item1.zst (71/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample63-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 351, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 351, 128) -Output shape: (1, 351, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.output: torch.Size([1, 351, 4096]) -> torch.Size([1, 1, 351, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,148B, BPFP=0.2036 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,100B, BPFP=1.3154 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,032B, BPFP=0.6239 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,272B, BPFP=1.4083 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,632B, BPFP=0.8153 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,592B, BPFP=1.4154 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,004B, BPFP=0.8681 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,780B, BPFP=1.3751 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,404B, BPFP=0.6100 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,628B, BPFP=1.4385 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 143,388B, BPFP=0.7979 -⌛️ [2/4] FRONTEND: Frontend time: 0.405s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.output: torch.Size([1, 351, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.696s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.output: torch.Size([1, 351, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02435836 17.83013099 - layer.0.v_cache 0.00000027 0.00030740 - layer.1.k_cache 0.00294291 1.17008959 - layer.1.v_cache 0.00000079 0.00116632 - layer.2.k_cache 0.00116231 0.69427364 - layer.2.v_cache 0.00000118 0.00167509 - layer.3.k_cache 0.00133045 0.75837547 - layer.3.v_cache 0.00000222 0.00275877 - layer.4.k_cache 0.00357138 1.38104100 - layer.4.v_cache 0.00000341 0.00462612 - layer.4.output 0.00014419 0.08697617 - ------------------------------------------------------------------------------------- - TOTAL 0.00242500 1.58516779 - (elements=5,031,936) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5031936 -Total Bytes 595980 -BPFP 0.9475 bits/point -EBPFP 1.8950 equivalent bits/point -MSE 1.585168 ----------------------- -------------------------------------------------------- -Time: 1.119s Load: 0.018s, Pack+Encode: 0.405s, Decode+Unpack: 0.696s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 351, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5852 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample63-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample63-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample64-layer4-item1.zst (72/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample64-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 323, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 323, 128) -Output shape: (1, 323, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.output: torch.Size([1, 323, 4096]) -> torch.Size([1, 1, 323, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,604B, BPFP=0.2081 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 55,604B, BPFP=1.3449 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,120B, BPFP=0.6560 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 59,824B, BPFP=1.4470 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,744B, BPFP=0.8404 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 60,972B, BPFP=1.4747 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,536B, BPFP=0.9079 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 59,204B, BPFP=1.4320 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,392B, BPFP=0.6384 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 61,352B, BPFP=1.4839 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 138,412B, BPFP=0.8370 -⌛️ [2/4] FRONTEND: Frontend time: 0.412s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.output: torch.Size([1, 323, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.689s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.output: torch.Size([1, 323, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02343196 18.86923435 - layer.0.v_cache 0.00000026 0.00030489 - layer.1.k_cache 0.00287105 1.21561217 - layer.1.v_cache 0.00000078 0.00116071 - layer.2.k_cache 0.00117289 0.71936101 - layer.2.v_cache 0.00000112 0.00168206 - layer.3.k_cache 0.00130124 0.76539824 - layer.3.v_cache 0.00000213 0.00268020 - layer.4.k_cache 0.00357998 1.42131633 - layer.4.v_cache 0.00000342 0.00463861 - layer.4.output 0.00014651 0.09318703 - ------------------------------------------------------------------------------------- - TOTAL 0.00235363 1.66958119 - (elements=4,630,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4630528 -Total Bytes 569764 -BPFP 0.9844 bits/point -EBPFP 1.9687 equivalent bits/point -MSE 1.669581 ----------------------- -------------------------------------------------------- -Time: 1.119s Load: 0.018s, Pack+Encode: 0.412s, Decode+Unpack: 0.689s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 323, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6696 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample64-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample64-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample65-layer4-item1.zst (73/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample65-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 356, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.020s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 356, 128) -Output shape: (1, 356, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.0.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.1.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.1.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.2.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.2.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.3.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.3.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.4.k_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.4.v_cache: torch.Size([1, 8, 356, 128]) -> torch.Size([1, 1, 356, 1024]) - layer.4.output: torch.Size([1, 356, 4096]) -> torch.Size([1, 1, 356, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,096B, BPFP=0.1996 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,860B, BPFP=1.3136 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,864B, BPFP=0.6334 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,620B, BPFP=1.3962 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,304B, BPFP=0.8186 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,324B, BPFP=1.4116 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,252B, BPFP=0.8833 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,580B, BPFP=1.3733 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,464B, BPFP=0.6027 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,216B, BPFP=1.4312 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 147,716B, BPFP=0.8104 -⌛️ [2/4] FRONTEND: Frontend time: 0.409s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 356, 128]) - layer.0.v_cache: torch.Size([1, 8, 356, 128]) - layer.1.k_cache: torch.Size([1, 8, 356, 128]) - layer.1.v_cache: torch.Size([1, 8, 356, 128]) - layer.2.k_cache: torch.Size([1, 8, 356, 128]) - layer.2.v_cache: torch.Size([1, 8, 356, 128]) - layer.3.k_cache: torch.Size([1, 8, 356, 128]) - layer.3.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.k_cache: torch.Size([1, 8, 356, 128]) - layer.4.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.output: torch.Size([1, 356, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.701s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 356, 128]) - layer.0.v_cache: torch.Size([1, 8, 356, 128]) - layer.1.k_cache: torch.Size([1, 8, 356, 128]) - layer.1.v_cache: torch.Size([1, 8, 356, 128]) - layer.2.k_cache: torch.Size([1, 8, 356, 128]) - layer.2.v_cache: torch.Size([1, 8, 356, 128]) - layer.3.k_cache: torch.Size([1, 8, 356, 128]) - layer.3.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.k_cache: torch.Size([1, 8, 356, 128]) - layer.4.v_cache: torch.Size([1, 8, 356, 128]) - layer.4.output: torch.Size([1, 356, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02477347 18.90143988 - layer.0.v_cache 0.00000027 0.00031724 - layer.1.k_cache 0.00294478 1.28124083 - layer.1.v_cache 0.00000080 0.00117353 - layer.2.k_cache 0.00115317 0.69590048 - layer.2.v_cache 0.00000113 0.00169472 - layer.3.k_cache 0.00130758 0.76793756 - layer.3.v_cache 0.00000212 0.00277080 - layer.4.k_cache 0.00357902 1.40664381 - layer.4.v_cache 0.00000333 0.00474114 - layer.4.output 0.00014592 0.09242284 - ------------------------------------------------------------------------------------- - TOTAL 0.00245352 1.67382510 - (elements=5,103,616) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5103616 -Total Bytes 606296 -BPFP 0.9504 bits/point -EBPFP 1.9008 equivalent bits/point -MSE 1.673825 ----------------------- -------------------------------------------------------- -Time: 1.131s Load: 0.020s, Pack+Encode: 0.409s, Decode+Unpack: 0.701s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 356, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 356, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6738 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample65-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample65-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample66-layer4-item1.zst (74/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample66-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 342, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 342, 128) -Output shape: (1, 342, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) -> torch.Size([1, 1, 342, 1024]) - layer.4.output: torch.Size([1, 342, 4096]) -> torch.Size([1, 1, 342, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,048B, BPFP=0.2067 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,416B, BPFP=1.3573 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,604B, BPFP=0.6306 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,180B, BPFP=1.4661 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,824B, BPFP=0.8412 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,212B, BPFP=1.4897 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,424B, BPFP=0.9006 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,644B, BPFP=1.4539 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,012B, BPFP=0.6171 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,740B, BPFP=1.5017 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,244B, BPFP=0.8295 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.output: torch.Size([1, 342, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 342, 128]) - layer.0.v_cache: torch.Size([1, 8, 342, 128]) - layer.1.k_cache: torch.Size([1, 8, 342, 128]) - layer.1.v_cache: torch.Size([1, 8, 342, 128]) - layer.2.k_cache: torch.Size([1, 8, 342, 128]) - layer.2.v_cache: torch.Size([1, 8, 342, 128]) - layer.3.k_cache: torch.Size([1, 8, 342, 128]) - layer.3.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.k_cache: torch.Size([1, 8, 342, 128]) - layer.4.v_cache: torch.Size([1, 8, 342, 128]) - layer.4.output: torch.Size([1, 342, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02260964 18.59486933 - layer.0.v_cache 0.00000027 0.00031290 - layer.1.k_cache 0.00288942 1.22696753 - layer.1.v_cache 0.00000084 0.00116522 - layer.2.k_cache 0.00116702 0.70083092 - layer.2.v_cache 0.00000122 0.00169237 - layer.3.k_cache 0.00131477 0.75431547 - layer.3.v_cache 0.00000216 0.00274234 - layer.4.k_cache 0.00350203 1.37436011 - layer.4.v_cache 0.00000356 0.00474522 - layer.4.output 0.00013713 0.09330028 - ------------------------------------------------------------------------------------- - TOTAL 0.00228853 1.64537161 - (elements=4,902,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4902912 -Total Bytes 603348 -BPFP 0.9845 bits/point -EBPFP 1.9689 equivalent bits/point -MSE 1.645372 ----------------------- -------------------------------------------------------- -Time: 1.105s Load: 0.019s, Pack+Encode: 0.408s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 342, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 342, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6454 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample66-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample66-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample67-layer4-item1.zst (75/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample67-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 360, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.020s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 360, 128) -Output shape: (1, 360, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.0.v_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.1.k_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.1.v_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.2.k_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.2.v_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.3.k_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.3.v_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.4.k_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.4.v_cache: torch.Size([1, 8, 360, 128]) -> torch.Size([1, 1, 360, 1024]) - layer.4.output: torch.Size([1, 360, 4096]) -> torch.Size([1, 1, 360, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,148B, BPFP=0.1985 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 60,600B, BPFP=1.3151 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,424B, BPFP=0.6385 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,212B, BPFP=1.3935 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 37,768B, BPFP=0.8196 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,976B, BPFP=1.4101 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 40,484B, BPFP=0.8786 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,644B, BPFP=1.3812 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,892B, BPFP=0.6270 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,292B, BPFP=1.4386 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 149,688B, BPFP=0.8121 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 360, 128]) - layer.0.v_cache: torch.Size([1, 8, 360, 128]) - layer.1.k_cache: torch.Size([1, 8, 360, 128]) - layer.1.v_cache: torch.Size([1, 8, 360, 128]) - layer.2.k_cache: torch.Size([1, 8, 360, 128]) - layer.2.v_cache: torch.Size([1, 8, 360, 128]) - layer.3.k_cache: torch.Size([1, 8, 360, 128]) - layer.3.v_cache: torch.Size([1, 8, 360, 128]) - layer.4.k_cache: torch.Size([1, 8, 360, 128]) - layer.4.v_cache: torch.Size([1, 8, 360, 128]) - layer.4.output: torch.Size([1, 360, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.685s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 360, 128]) - layer.0.v_cache: torch.Size([1, 8, 360, 128]) - layer.1.k_cache: torch.Size([1, 8, 360, 128]) - layer.1.v_cache: torch.Size([1, 8, 360, 128]) - layer.2.k_cache: torch.Size([1, 8, 360, 128]) - layer.2.v_cache: torch.Size([1, 8, 360, 128]) - layer.3.k_cache: torch.Size([1, 8, 360, 128]) - layer.3.v_cache: torch.Size([1, 8, 360, 128]) - layer.4.k_cache: torch.Size([1, 8, 360, 128]) - layer.4.v_cache: torch.Size([1, 8, 360, 128]) - layer.4.output: torch.Size([1, 360, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02342221 18.77050239 - layer.0.v_cache 0.00000028 0.00031657 - layer.1.k_cache 0.00285409 1.17796232 - layer.1.v_cache 0.00000079 0.00117710 - layer.2.k_cache 0.00117169 0.69275055 - layer.2.v_cache 0.00000113 0.00170143 - layer.3.k_cache 0.00130395 0.76656791 - layer.3.v_cache 0.00000220 0.00286073 - layer.4.k_cache 0.00353751 1.39290407 - layer.4.v_cache 0.00000341 0.00488297 - layer.4.output 0.00014250 0.09262073 - ------------------------------------------------------------------------------------- - TOTAL 0.00234766 1.65586493 - (elements=5,160,960) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5160960 -Total Bytes 615128 -BPFP 0.9535 bits/point -EBPFP 1.9070 equivalent bits/point -MSE 1.655865 ----------------------- -------------------------------------------------------- -Time: 1.112s Load: 0.020s, Pack+Encode: 0.407s, Decode+Unpack: 0.685s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 360, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 360, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6559 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample67-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample67-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample68-layer4-item1.zst (76/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample68-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 294, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.015s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 294, 128) -Output shape: (1, 294, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.0.v_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.1.k_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.1.v_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.2.k_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.2.v_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.3.k_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.3.v_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.4.k_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.4.v_cache: torch.Size([1, 8, 294, 128]) -> torch.Size([1, 1, 294, 1024]) - layer.4.output: torch.Size([1, 294, 4096]) -> torch.Size([1, 1, 294, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,656B, BPFP=0.2034 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,084B, BPFP=1.3309 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,388B, BPFP=0.6481 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,380B, BPFP=1.4185 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,028B, BPFP=0.8245 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 54,224B, BPFP=1.4409 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,104B, BPFP=0.8797 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 52,784B, BPFP=1.4026 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,548B, BPFP=0.6257 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,060B, BPFP=1.4631 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 123,776B, BPFP=0.8223 -⌛️ [2/4] FRONTEND: Frontend time: 0.357s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 294, 128]) - layer.0.v_cache: torch.Size([1, 8, 294, 128]) - layer.1.k_cache: torch.Size([1, 8, 294, 128]) - layer.1.v_cache: torch.Size([1, 8, 294, 128]) - layer.2.k_cache: torch.Size([1, 8, 294, 128]) - layer.2.v_cache: torch.Size([1, 8, 294, 128]) - layer.3.k_cache: torch.Size([1, 8, 294, 128]) - layer.3.v_cache: torch.Size([1, 8, 294, 128]) - layer.4.k_cache: torch.Size([1, 8, 294, 128]) - layer.4.v_cache: torch.Size([1, 8, 294, 128]) - layer.4.output: torch.Size([1, 294, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.581s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 294, 128]) - layer.0.v_cache: torch.Size([1, 8, 294, 128]) - layer.1.k_cache: torch.Size([1, 8, 294, 128]) - layer.1.v_cache: torch.Size([1, 8, 294, 128]) - layer.2.k_cache: torch.Size([1, 8, 294, 128]) - layer.2.v_cache: torch.Size([1, 8, 294, 128]) - layer.3.k_cache: torch.Size([1, 8, 294, 128]) - layer.3.v_cache: torch.Size([1, 8, 294, 128]) - layer.4.k_cache: torch.Size([1, 8, 294, 128]) - layer.4.v_cache: torch.Size([1, 8, 294, 128]) - layer.4.output: torch.Size([1, 294, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02439844 19.88511772 - layer.0.v_cache 0.00000027 0.00030394 - layer.1.k_cache 0.00289989 1.22193172 - layer.1.v_cache 0.00000083 0.00112524 - layer.2.k_cache 0.00119190 0.69728197 - layer.2.v_cache 0.00000115 0.00168344 - layer.3.k_cache 0.00131407 0.76117867 - layer.3.v_cache 0.00000220 0.00276881 - layer.4.k_cache 0.00351532 1.38581589 - layer.4.v_cache 0.00000333 0.00471008 - layer.4.output 0.00015801 0.10103371 - ------------------------------------------------------------------------------------- - TOTAL 0.00242567 1.74043231 - (elements=4,214,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4214784 -Total Bytes 509032 -BPFP 0.9662 bits/point -EBPFP 1.9324 equivalent bits/point -MSE 1.740432 ----------------------- -------------------------------------------------------- -Time: 0.952s Load: 0.015s, Pack+Encode: 0.357s, Decode+Unpack: 0.581s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 294, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 294, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7404 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample68-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample68-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample69-layer4-item1.zst (77/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample69-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 338, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 338, 128) -Output shape: (1, 338, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.0.v_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.1.k_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.1.v_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.2.k_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.2.v_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.3.k_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.3.v_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.4.k_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.4.v_cache: torch.Size([1, 8, 338, 128]) -> torch.Size([1, 1, 338, 1024]) - layer.4.output: torch.Size([1, 338, 4096]) -> torch.Size([1, 1, 338, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,184B, BPFP=0.2123 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,544B, BPFP=1.3532 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,060B, BPFP=0.6486 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,652B, BPFP=1.4712 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,476B, BPFP=0.8431 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,920B, BPFP=1.5006 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,164B, BPFP=0.9052 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,312B, BPFP=1.4634 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,068B, BPFP=0.6256 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,684B, BPFP=1.5182 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 150,608B, BPFP=0.8703 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 338, 128]) - layer.0.v_cache: torch.Size([1, 8, 338, 128]) - layer.1.k_cache: torch.Size([1, 8, 338, 128]) - layer.1.v_cache: torch.Size([1, 8, 338, 128]) - layer.2.k_cache: torch.Size([1, 8, 338, 128]) - layer.2.v_cache: torch.Size([1, 8, 338, 128]) - layer.3.k_cache: torch.Size([1, 8, 338, 128]) - layer.3.v_cache: torch.Size([1, 8, 338, 128]) - layer.4.k_cache: torch.Size([1, 8, 338, 128]) - layer.4.v_cache: torch.Size([1, 8, 338, 128]) - layer.4.output: torch.Size([1, 338, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.680s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 338, 128]) - layer.0.v_cache: torch.Size([1, 8, 338, 128]) - layer.1.k_cache: torch.Size([1, 8, 338, 128]) - layer.1.v_cache: torch.Size([1, 8, 338, 128]) - layer.2.k_cache: torch.Size([1, 8, 338, 128]) - layer.2.v_cache: torch.Size([1, 8, 338, 128]) - layer.3.k_cache: torch.Size([1, 8, 338, 128]) - layer.3.v_cache: torch.Size([1, 8, 338, 128]) - layer.4.k_cache: torch.Size([1, 8, 338, 128]) - layer.4.v_cache: torch.Size([1, 8, 338, 128]) - layer.4.output: torch.Size([1, 338, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02370679 18.90583395 - layer.0.v_cache 0.00000026 0.00030736 - layer.1.k_cache 0.00288373 1.25446794 - layer.1.v_cache 0.00000083 0.00114492 - layer.2.k_cache 0.00117924 0.71019248 - layer.2.v_cache 0.00000113 0.00169323 - layer.3.k_cache 0.00131114 0.77584866 - layer.3.v_cache 0.00000214 0.00279369 - layer.4.k_cache 0.00355340 1.43691385 - layer.4.v_cache 0.00000355 0.00471723 - layer.4.output 0.00015048 0.09348430 - ------------------------------------------------------------------------------------- - TOTAL 0.00237458 1.67627503 - (elements=4,845,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4845568 -Total Bytes 606672 -BPFP 1.0016 bits/point -EBPFP 2.0032 equivalent bits/point -MSE 1.676275 ----------------------- -------------------------------------------------------- -Time: 1.106s Load: 0.018s, Pack+Encode: 0.408s, Decode+Unpack: 0.680s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 338, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 338, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6763 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample69-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample69-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample7-layer4-item1.zst (78/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample7-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 340, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 340, 128) -Output shape: (1, 340, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.0.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.1.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.1.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.2.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.2.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.3.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.3.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.4.k_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.4.v_cache: torch.Size([1, 8, 340, 128]) -> torch.Size([1, 1, 340, 1024]) - layer.4.output: torch.Size([1, 340, 4096]) -> torch.Size([1, 1, 340, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,268B, BPFP=0.2130 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,900B, BPFP=1.3534 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,124B, BPFP=0.6462 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,624B, BPFP=1.4619 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,408B, BPFP=0.8366 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 64,676B, BPFP=1.4861 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,356B, BPFP=0.9043 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,128B, BPFP=1.4506 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,264B, BPFP=0.6265 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,936B, BPFP=1.5151 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,944B, BPFP=0.8556 -⌛️ [2/4] FRONTEND: Frontend time: 0.411s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 340, 128]) - layer.0.v_cache: torch.Size([1, 8, 340, 128]) - layer.1.k_cache: torch.Size([1, 8, 340, 128]) - layer.1.v_cache: torch.Size([1, 8, 340, 128]) - layer.2.k_cache: torch.Size([1, 8, 340, 128]) - layer.2.v_cache: torch.Size([1, 8, 340, 128]) - layer.3.k_cache: torch.Size([1, 8, 340, 128]) - layer.3.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.k_cache: torch.Size([1, 8, 340, 128]) - layer.4.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.output: torch.Size([1, 340, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.682s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 340, 128]) - layer.0.v_cache: torch.Size([1, 8, 340, 128]) - layer.1.k_cache: torch.Size([1, 8, 340, 128]) - layer.1.v_cache: torch.Size([1, 8, 340, 128]) - layer.2.k_cache: torch.Size([1, 8, 340, 128]) - layer.2.v_cache: torch.Size([1, 8, 340, 128]) - layer.3.k_cache: torch.Size([1, 8, 340, 128]) - layer.3.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.k_cache: torch.Size([1, 8, 340, 128]) - layer.4.v_cache: torch.Size([1, 8, 340, 128]) - layer.4.output: torch.Size([1, 340, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02367318 18.80162856 - layer.0.v_cache 0.00000026 0.00031052 - layer.1.k_cache 0.00294777 1.21906343 - layer.1.v_cache 0.00000081 0.00116473 - layer.2.k_cache 0.00121079 0.70795216 - layer.2.v_cache 0.00000125 0.00171565 - layer.3.k_cache 0.00131075 0.77041572 - layer.3.v_cache 0.00000218 0.00283158 - layer.4.k_cache 0.00357831 1.39631312 - layer.4.v_cache 0.00000337 0.00485142 - layer.4.output 0.00014951 0.09840472 - ------------------------------------------------------------------------------------- - TOTAL 0.00238048 1.66427613 - (elements=4,874,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4874240 -Total Bytes 605628 -BPFP 0.9940 bits/point -EBPFP 1.9880 equivalent bits/point -MSE 1.664276 ----------------------- -------------------------------------------------------- -Time: 1.111s Load: 0.017s, Pack+Encode: 0.411s, Decode+Unpack: 0.682s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 340, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 340, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6643 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample7-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample7-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample70-layer4-item1.zst (79/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample70-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 320, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 320, 128) -Output shape: (1, 320, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.0.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.1.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.1.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.2.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.2.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.3.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.3.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.4.k_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.4.v_cache: torch.Size([1, 8, 320, 128]) -> torch.Size([1, 1, 320, 1024]) - layer.4.output: torch.Size([1, 320, 4096]) -> torch.Size([1, 1, 320, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,612B, BPFP=0.1858 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 49,212B, BPFP=1.2015 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,900B, BPFP=0.5591 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,184B, BPFP=1.2984 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,944B, BPFP=0.7555 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 54,352B, BPFP=1.3270 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,836B, BPFP=0.8017 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 52,984B, BPFP=1.2936 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,444B, BPFP=0.5479 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 54,968B, BPFP=1.3420 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 112,168B, BPFP=0.6846 -⌛️ [2/4] FRONTEND: Frontend time: 0.356s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 320, 128]) - layer.0.v_cache: torch.Size([1, 8, 320, 128]) - layer.1.k_cache: torch.Size([1, 8, 320, 128]) - layer.1.v_cache: torch.Size([1, 8, 320, 128]) - layer.2.k_cache: torch.Size([1, 8, 320, 128]) - layer.2.v_cache: torch.Size([1, 8, 320, 128]) - layer.3.k_cache: torch.Size([1, 8, 320, 128]) - layer.3.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.k_cache: torch.Size([1, 8, 320, 128]) - layer.4.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.output: torch.Size([1, 320, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.582s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 320, 128]) - layer.0.v_cache: torch.Size([1, 8, 320, 128]) - layer.1.k_cache: torch.Size([1, 8, 320, 128]) - layer.1.v_cache: torch.Size([1, 8, 320, 128]) - layer.2.k_cache: torch.Size([1, 8, 320, 128]) - layer.2.v_cache: torch.Size([1, 8, 320, 128]) - layer.3.k_cache: torch.Size([1, 8, 320, 128]) - layer.3.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.k_cache: torch.Size([1, 8, 320, 128]) - layer.4.v_cache: torch.Size([1, 8, 320, 128]) - layer.4.output: torch.Size([1, 320, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02419060 17.32916870 - layer.0.v_cache 0.00000027 0.00029273 - layer.1.k_cache 0.00288080 1.01749773 - layer.1.v_cache 0.00000084 0.00114040 - layer.2.k_cache 0.00117023 0.65717411 - layer.2.v_cache 0.00000114 0.00164237 - layer.3.k_cache 0.00131996 0.73167510 - layer.3.v_cache 0.00000215 0.00265052 - layer.4.k_cache 0.00348745 1.28000107 - layer.4.v_cache 0.00000346 0.00454543 - layer.4.output 0.00014756 0.08617021 - ------------------------------------------------------------------------------------- - TOTAL 0.00240337 1.52646207 - (elements=4,587,520) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4587520 -Total Bytes 493604 -BPFP 0.8608 bits/point -EBPFP 1.7216 equivalent bits/point -MSE 1.526462 ----------------------- -------------------------------------------------------- -Time: 0.956s Load: 0.017s, Pack+Encode: 0.356s, Decode+Unpack: 0.582s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 320, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 320, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5265 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample70-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample70-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample71-layer4-item1.zst (80/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample71-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 328, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 328, 128) -Output shape: (1, 328, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.0.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.1.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.1.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.2.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.2.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.3.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.3.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.4.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.4.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.4.output: torch.Size([1, 328, 4096]) -> torch.Size([1, 1, 328, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,888B, BPFP=0.2117 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,224B, BPFP=1.3630 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,684B, BPFP=0.6594 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,672B, BPFP=1.4689 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,888B, BPFP=0.8310 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,144B, BPFP=1.5040 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,388B, BPFP=0.9143 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,388B, BPFP=1.4622 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,000B, BPFP=0.6431 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,760B, BPFP=1.5187 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,916B, BPFP=0.8629 -⌛️ [2/4] FRONTEND: Frontend time: 0.418s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 328, 128]) - layer.0.v_cache: torch.Size([1, 8, 328, 128]) - layer.1.k_cache: torch.Size([1, 8, 328, 128]) - layer.1.v_cache: torch.Size([1, 8, 328, 128]) - layer.2.k_cache: torch.Size([1, 8, 328, 128]) - layer.2.v_cache: torch.Size([1, 8, 328, 128]) - layer.3.k_cache: torch.Size([1, 8, 328, 128]) - layer.3.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.k_cache: torch.Size([1, 8, 328, 128]) - layer.4.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.output: torch.Size([1, 328, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.687s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 328, 128]) - layer.0.v_cache: torch.Size([1, 8, 328, 128]) - layer.1.k_cache: torch.Size([1, 8, 328, 128]) - layer.1.v_cache: torch.Size([1, 8, 328, 128]) - layer.2.k_cache: torch.Size([1, 8, 328, 128]) - layer.2.v_cache: torch.Size([1, 8, 328, 128]) - layer.3.k_cache: torch.Size([1, 8, 328, 128]) - layer.3.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.k_cache: torch.Size([1, 8, 328, 128]) - layer.4.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.output: torch.Size([1, 328, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02411815 20.88441728 - layer.0.v_cache 0.00000026 0.00031055 - layer.1.k_cache 0.00293126 1.24764205 - layer.1.v_cache 0.00000079 0.00114264 - layer.2.k_cache 0.00115822 0.68766478 - layer.2.v_cache 0.00000116 0.00166568 - layer.3.k_cache 0.00132993 0.76124950 - layer.3.v_cache 0.00000214 0.00270448 - layer.4.k_cache 0.00358323 1.43185499 - layer.4.v_cache 0.00000336 0.00459026 - layer.4.output 0.00014415 0.09399020 - ------------------------------------------------------------------------------------- - TOTAL 0.00240751 1.81422879 - (elements=4,702,208) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4702208 -Total Bytes 588952 -BPFP 1.0020 bits/point -EBPFP 2.0040 equivalent bits/point -MSE 1.814229 ----------------------- -------------------------------------------------------- -Time: 1.122s Load: 0.018s, Pack+Encode: 0.418s, Decode+Unpack: 0.687s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 328, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.8142 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample71-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample71-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample72-layer4-item1.zst (81/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample72-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 328, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 328, 128) -Output shape: (1, 328, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.0.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.1.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.1.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.2.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.2.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.3.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.3.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.4.k_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.4.v_cache: torch.Size([1, 8, 328, 128]) -> torch.Size([1, 1, 328, 1024]) - layer.4.output: torch.Size([1, 328, 4096]) -> torch.Size([1, 1, 328, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,928B, BPFP=0.2127 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,868B, BPFP=1.3545 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,336B, BPFP=0.6511 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,720B, BPFP=1.4701 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,148B, BPFP=0.8372 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,168B, BPFP=1.5046 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,220B, BPFP=0.9103 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,448B, BPFP=1.4636 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,264B, BPFP=0.6256 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,544B, BPFP=1.5135 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 140,892B, BPFP=0.8390 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 328, 128]) - layer.0.v_cache: torch.Size([1, 8, 328, 128]) - layer.1.k_cache: torch.Size([1, 8, 328, 128]) - layer.1.v_cache: torch.Size([1, 8, 328, 128]) - layer.2.k_cache: torch.Size([1, 8, 328, 128]) - layer.2.v_cache: torch.Size([1, 8, 328, 128]) - layer.3.k_cache: torch.Size([1, 8, 328, 128]) - layer.3.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.k_cache: torch.Size([1, 8, 328, 128]) - layer.4.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.output: torch.Size([1, 328, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.692s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 328, 128]) - layer.0.v_cache: torch.Size([1, 8, 328, 128]) - layer.1.k_cache: torch.Size([1, 8, 328, 128]) - layer.1.v_cache: torch.Size([1, 8, 328, 128]) - layer.2.k_cache: torch.Size([1, 8, 328, 128]) - layer.2.v_cache: torch.Size([1, 8, 328, 128]) - layer.3.k_cache: torch.Size([1, 8, 328, 128]) - layer.3.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.k_cache: torch.Size([1, 8, 328, 128]) - layer.4.v_cache: torch.Size([1, 8, 328, 128]) - layer.4.output: torch.Size([1, 328, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02582944 19.33510087 - layer.0.v_cache 0.00000026 0.00030886 - layer.1.k_cache 0.00293683 1.21156469 - layer.1.v_cache 0.00000081 0.00116555 - layer.2.k_cache 0.00118071 0.70122998 - layer.2.v_cache 0.00000117 0.00172001 - layer.3.k_cache 0.00133336 0.76824602 - layer.3.v_cache 0.00000214 0.00269633 - layer.4.k_cache 0.00361319 1.41337837 - layer.4.v_cache 0.00000321 0.00448697 - layer.4.output 0.00015382 0.09357672 - ------------------------------------------------------------------------------------- - TOTAL 0.00253689 1.70101461 - (elements=4,702,208) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4702208 -Total Bytes 583536 -BPFP 0.9928 bits/point -EBPFP 1.9856 equivalent bits/point -MSE 1.701015 ----------------------- -------------------------------------------------------- -Time: 1.117s Load: 0.017s, Pack+Encode: 0.408s, Decode+Unpack: 0.692s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 328, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 328, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7010 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample72-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample72-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample73-layer4-item1.zst (82/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample73-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 352, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 352, 128) -Output shape: (1, 352, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.0.v_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.1.k_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.1.v_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.2.k_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.2.v_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.3.k_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.3.v_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.4.k_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.4.v_cache: torch.Size([1, 8, 352, 128]) -> torch.Size([1, 1, 352, 1024]) - layer.4.output: torch.Size([1, 352, 4096]) -> torch.Size([1, 1, 352, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,864B, BPFP=0.1967 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,816B, BPFP=1.3054 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,592B, BPFP=0.6124 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,416B, BPFP=1.3853 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,864B, BPFP=0.7960 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,080B, BPFP=1.4000 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,804B, BPFP=0.8612 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,880B, BPFP=1.3734 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,844B, BPFP=0.5958 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,436B, BPFP=1.4301 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 135,632B, BPFP=0.7526 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 352, 128]) - layer.0.v_cache: torch.Size([1, 8, 352, 128]) - layer.1.k_cache: torch.Size([1, 8, 352, 128]) - layer.1.v_cache: torch.Size([1, 8, 352, 128]) - layer.2.k_cache: torch.Size([1, 8, 352, 128]) - layer.2.v_cache: torch.Size([1, 8, 352, 128]) - layer.3.k_cache: torch.Size([1, 8, 352, 128]) - layer.3.v_cache: torch.Size([1, 8, 352, 128]) - layer.4.k_cache: torch.Size([1, 8, 352, 128]) - layer.4.v_cache: torch.Size([1, 8, 352, 128]) - layer.4.output: torch.Size([1, 352, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.673s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 352, 128]) - layer.0.v_cache: torch.Size([1, 8, 352, 128]) - layer.1.k_cache: torch.Size([1, 8, 352, 128]) - layer.1.v_cache: torch.Size([1, 8, 352, 128]) - layer.2.k_cache: torch.Size([1, 8, 352, 128]) - layer.2.v_cache: torch.Size([1, 8, 352, 128]) - layer.3.k_cache: torch.Size([1, 8, 352, 128]) - layer.3.v_cache: torch.Size([1, 8, 352, 128]) - layer.4.k_cache: torch.Size([1, 8, 352, 128]) - layer.4.v_cache: torch.Size([1, 8, 352, 128]) - layer.4.output: torch.Size([1, 352, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02395736 18.19969593 - layer.0.v_cache 0.00000026 0.00030655 - layer.1.k_cache 0.00288394 1.24271705 - layer.1.v_cache 0.00000077 0.00115179 - layer.2.k_cache 0.00114480 0.69133642 - layer.2.v_cache 0.00000114 0.00172018 - layer.3.k_cache 0.00130173 0.75724272 - layer.3.v_cache 0.00000215 0.00285071 - layer.4.k_cache 0.00352615 1.37152351 - layer.4.v_cache 0.00000336 0.00486789 - layer.4.output 0.00014232 0.08799266 - ------------------------------------------------------------------------------------- - TOTAL 0.00238507 1.61609881 - (elements=5,046,272) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5046272 -Total Bytes 584228 -BPFP 0.9262 bits/point -EBPFP 1.8524 equivalent bits/point -MSE 1.616099 ----------------------- -------------------------------------------------------- -Time: 1.098s Load: 0.019s, Pack+Encode: 0.407s, Decode+Unpack: 0.673s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 352, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 352, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6161 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample73-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample73-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample74-layer4-item1.zst (83/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample74-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 323, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 323, 128) -Output shape: (1, 323, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.output: torch.Size([1, 323, 4096]) -> torch.Size([1, 1, 323, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,864B, BPFP=0.2144 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 55,980B, BPFP=1.3540 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,396B, BPFP=0.6626 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 59,840B, BPFP=1.4474 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,464B, BPFP=0.8578 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 61,508B, BPFP=1.4877 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,784B, BPFP=0.9139 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 59,800B, BPFP=1.4464 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,092B, BPFP=0.6553 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 62,060B, BPFP=1.5011 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 151,744B, BPFP=0.9176 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.output: torch.Size([1, 323, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.673s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.output: torch.Size([1, 323, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02435790 18.92503598 - layer.0.v_cache 0.00000027 0.00030878 - layer.1.k_cache 0.00290123 1.21783353 - layer.1.v_cache 0.00000082 0.00116804 - layer.2.k_cache 0.00116542 0.69674376 - layer.2.v_cache 0.00000118 0.00173493 - layer.3.k_cache 0.00132756 0.76743343 - layer.3.v_cache 0.00000224 0.00287054 - layer.4.k_cache 0.00351721 1.39002438 - layer.4.v_cache 0.00000335 0.00482269 - layer.4.output 0.00015064 0.09654518 - ------------------------------------------------------------------------------------- - TOTAL 0.00241998 1.67101120 - (elements=4,630,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4630528 -Total Bytes 587532 -BPFP 1.0151 bits/point -EBPFP 2.0301 equivalent bits/point -MSE 1.671011 ----------------------- -------------------------------------------------------- -Time: 1.096s Load: 0.017s, Pack+Encode: 0.406s, Decode+Unpack: 0.673s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 323, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6710 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample74-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample74-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample75-layer4-item1.zst (84/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample75-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 327, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 327, 128) -Output shape: (1, 327, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.0.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.1.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.1.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.2.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.2.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.3.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.3.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.4.k_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.4.v_cache: torch.Size([1, 8, 327, 128]) -> torch.Size([1, 1, 327, 1024]) - layer.4.output: torch.Size([1, 327, 4096]) -> torch.Size([1, 1, 327, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,876B, BPFP=0.2121 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,956B, BPFP=1.3608 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,300B, BPFP=0.6761 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,708B, BPFP=1.4743 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,452B, BPFP=0.8470 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,148B, BPFP=1.5087 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,308B, BPFP=0.9152 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,468B, BPFP=1.4686 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,348B, BPFP=0.6534 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,232B, BPFP=1.5107 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 147,060B, BPFP=0.8784 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 327, 128]) - layer.0.v_cache: torch.Size([1, 8, 327, 128]) - layer.1.k_cache: torch.Size([1, 8, 327, 128]) - layer.1.v_cache: torch.Size([1, 8, 327, 128]) - layer.2.k_cache: torch.Size([1, 8, 327, 128]) - layer.2.v_cache: torch.Size([1, 8, 327, 128]) - layer.3.k_cache: torch.Size([1, 8, 327, 128]) - layer.3.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.k_cache: torch.Size([1, 8, 327, 128]) - layer.4.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.output: torch.Size([1, 327, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.674s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 327, 128]) - layer.0.v_cache: torch.Size([1, 8, 327, 128]) - layer.1.k_cache: torch.Size([1, 8, 327, 128]) - layer.1.v_cache: torch.Size([1, 8, 327, 128]) - layer.2.k_cache: torch.Size([1, 8, 327, 128]) - layer.2.v_cache: torch.Size([1, 8, 327, 128]) - layer.3.k_cache: torch.Size([1, 8, 327, 128]) - layer.3.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.k_cache: torch.Size([1, 8, 327, 128]) - layer.4.v_cache: torch.Size([1, 8, 327, 128]) - layer.4.output: torch.Size([1, 327, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02420299 20.14289618 - layer.0.v_cache 0.00000027 0.00030493 - layer.1.k_cache 0.00291079 1.23886248 - layer.1.v_cache 0.00000081 0.00116447 - layer.2.k_cache 0.00118771 0.69288472 - layer.2.v_cache 0.00000120 0.00171636 - layer.3.k_cache 0.00132021 0.76409086 - layer.3.v_cache 0.00000219 0.00280721 - layer.4.k_cache 0.00351552 1.36998672 - layer.4.v_cache 0.00000327 0.00463802 - layer.4.output 0.00014986 0.09695039 - ------------------------------------------------------------------------------------- - TOTAL 0.00241031 1.75765382 - (elements=4,687,872) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4687872 -Total Bytes 591856 -BPFP 1.0100 bits/point -EBPFP 2.0200 equivalent bits/point -MSE 1.757654 ----------------------- -------------------------------------------------------- -Time: 1.094s Load: 0.017s, Pack+Encode: 0.402s, Decode+Unpack: 0.674s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 327, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 327, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7577 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample75-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample75-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample76-layer4-item1.zst (85/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample76-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 330, 128) -Output shape: (1, 330, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) -> torch.Size([1, 1, 330, 1024]) - layer.4.output: torch.Size([1, 330, 4096]) -> torch.Size([1, 1, 330, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,912B, BPFP=0.2110 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,160B, BPFP=1.3532 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,452B, BPFP=0.6736 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,292B, BPFP=1.4747 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,708B, BPFP=0.8454 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,924B, BPFP=1.5134 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,436B, BPFP=0.9099 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,256B, BPFP=1.4739 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,216B, BPFP=0.6443 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,264B, BPFP=1.5214 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 153,064B, BPFP=0.9059 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.694s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 330, 128]) - layer.0.v_cache: torch.Size([1, 8, 330, 128]) - layer.1.k_cache: torch.Size([1, 8, 330, 128]) - layer.1.v_cache: torch.Size([1, 8, 330, 128]) - layer.2.k_cache: torch.Size([1, 8, 330, 128]) - layer.2.v_cache: torch.Size([1, 8, 330, 128]) - layer.3.k_cache: torch.Size([1, 8, 330, 128]) - layer.3.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.k_cache: torch.Size([1, 8, 330, 128]) - layer.4.v_cache: torch.Size([1, 8, 330, 128]) - layer.4.output: torch.Size([1, 330, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02537568 20.29952799 - layer.0.v_cache 0.00000027 0.00030412 - layer.1.k_cache 0.00294190 1.23209016 - layer.1.v_cache 0.00000076 0.00109227 - layer.2.k_cache 0.00116180 0.69667257 - layer.2.v_cache 0.00000113 0.00160775 - layer.3.k_cache 0.00131353 0.75981154 - layer.3.v_cache 0.00000217 0.00266987 - layer.4.k_cache 0.00362690 1.45374146 - layer.4.v_cache 0.00000322 0.00451267 - layer.4.output 0.00015895 0.10041334 - ------------------------------------------------------------------------------------- - TOTAL 0.00250451 1.77526312 - (elements=4,730,880) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4730880 -Total Bytes 601684 -BPFP 1.0175 bits/point -EBPFP 2.0349 equivalent bits/point -MSE 1.775263 ----------------------- -------------------------------------------------------- -Time: 1.114s Load: 0.018s, Pack+Encode: 0.402s, Decode+Unpack: 0.694s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 330, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 330, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7753 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample76-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample76-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample77-layer4-item1.zst (86/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample77-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 329, 128) -Output shape: (1, 329, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.output: torch.Size([1, 329, 4096]) -> torch.Size([1, 1, 329, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,796B, BPFP=0.2089 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,632B, BPFP=1.3448 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,376B, BPFP=0.6501 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,020B, BPFP=1.4727 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,080B, BPFP=0.8330 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,208B, BPFP=1.5009 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,112B, BPFP=0.9050 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,428B, BPFP=1.4587 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,752B, BPFP=0.6353 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,596B, BPFP=1.5102 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,516B, BPFP=0.8401 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.681s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02321163 20.25253936 - layer.0.v_cache 0.00000026 0.00031125 - layer.1.k_cache 0.00292509 1.27882761 - layer.1.v_cache 0.00000078 0.00114418 - layer.2.k_cache 0.00118155 0.71770020 - layer.2.v_cache 0.00000112 0.00164742 - layer.3.k_cache 0.00133849 0.77406552 - layer.3.v_cache 0.00000208 0.00265644 - layer.4.k_cache 0.00360187 1.42383684 - layer.4.v_cache 0.00000304 0.00444178 - layer.4.output 0.00015018 0.09700798 - ------------------------------------------------------------------------------------- - TOTAL 0.00234762 1.77465732 - (elements=4,716,544) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4716544 -Total Bytes 584516 -BPFP 0.9914 bits/point -EBPFP 1.9829 equivalent bits/point -MSE 1.774657 ----------------------- -------------------------------------------------------- -Time: 1.104s Load: 0.017s, Pack+Encode: 0.406s, Decode+Unpack: 0.681s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7747 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample77-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample77-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample78-layer4-item1.zst (87/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample78-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 333, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 333, 128) -Output shape: (1, 333, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.output: torch.Size([1, 333, 4096]) -> torch.Size([1, 1, 333, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,160B, BPFP=0.2149 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,320B, BPFP=1.3448 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,360B, BPFP=0.6419 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,628B, BPFP=1.4693 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,920B, BPFP=0.8427 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,700B, BPFP=1.4945 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,732B, BPFP=0.9087 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,620B, BPFP=1.4691 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,116B, BPFP=0.6362 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 65,068B, BPFP=1.5266 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 146,312B, BPFP=0.8582 -⌛️ [2/4] FRONTEND: Frontend time: 0.402s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.output: torch.Size([1, 333, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.671s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.output: torch.Size([1, 333, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02306361 19.09888942 - layer.0.v_cache 0.00000026 0.00030467 - layer.1.k_cache 0.00286224 1.25490866 - layer.1.v_cache 0.00000078 0.00111791 - layer.2.k_cache 0.00116467 0.70671503 - layer.2.v_cache 0.00000115 0.00167333 - layer.3.k_cache 0.00135416 0.77072950 - layer.3.v_cache 0.00000234 0.00282759 - layer.4.k_cache 0.00356469 1.45364536 - layer.4.v_cache 0.00000325 0.00462008 - layer.4.output 0.00015110 0.09389022 - ------------------------------------------------------------------------------------- - TOTAL 0.00233011 1.69078517 - (elements=4,773,888) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4773888 -Total Bytes 595936 -BPFP 0.9987 bits/point -EBPFP 1.9973 equivalent bits/point -MSE 1.690785 ----------------------- -------------------------------------------------------- -Time: 1.090s Load: 0.017s, Pack+Encode: 0.402s, Decode+Unpack: 0.671s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 333, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6908 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample78-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample78-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample79-layer4-item1.zst (88/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample79-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 304, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.016s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 304, 128) -Output shape: (1, 304, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.0.v_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.1.k_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.1.v_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.2.k_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.2.v_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.3.k_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.3.v_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.4.k_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.4.v_cache: torch.Size([1, 8, 304, 128]) -> torch.Size([1, 1, 304, 1024]) - layer.4.output: torch.Size([1, 304, 4096]) -> torch.Size([1, 1, 304, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,700B, BPFP=0.1979 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 49,472B, BPFP=1.2714 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,564B, BPFP=0.6056 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,740B, BPFP=1.3811 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,844B, BPFP=0.7927 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 54,504B, BPFP=1.4007 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,572B, BPFP=0.8371 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 52,764B, BPFP=1.3560 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,980B, BPFP=0.5906 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 55,004B, BPFP=1.4135 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 131,260B, BPFP=0.8433 -⌛️ [2/4] FRONTEND: Frontend time: 0.348s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 304, 128]) - layer.0.v_cache: torch.Size([1, 8, 304, 128]) - layer.1.k_cache: torch.Size([1, 8, 304, 128]) - layer.1.v_cache: torch.Size([1, 8, 304, 128]) - layer.2.k_cache: torch.Size([1, 8, 304, 128]) - layer.2.v_cache: torch.Size([1, 8, 304, 128]) - layer.3.k_cache: torch.Size([1, 8, 304, 128]) - layer.3.v_cache: torch.Size([1, 8, 304, 128]) - layer.4.k_cache: torch.Size([1, 8, 304, 128]) - layer.4.v_cache: torch.Size([1, 8, 304, 128]) - layer.4.output: torch.Size([1, 304, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.573s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 304, 128]) - layer.0.v_cache: torch.Size([1, 8, 304, 128]) - layer.1.k_cache: torch.Size([1, 8, 304, 128]) - layer.1.v_cache: torch.Size([1, 8, 304, 128]) - layer.2.k_cache: torch.Size([1, 8, 304, 128]) - layer.2.v_cache: torch.Size([1, 8, 304, 128]) - layer.3.k_cache: torch.Size([1, 8, 304, 128]) - layer.3.v_cache: torch.Size([1, 8, 304, 128]) - layer.4.k_cache: torch.Size([1, 8, 304, 128]) - layer.4.v_cache: torch.Size([1, 8, 304, 128]) - layer.4.output: torch.Size([1, 304, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02416075 19.38056223 - layer.0.v_cache 0.00000027 0.00030175 - layer.1.k_cache 0.00290015 1.19137513 - layer.1.v_cache 0.00000081 0.00115071 - layer.2.k_cache 0.00116609 0.69321798 - layer.2.v_cache 0.00000115 0.00168309 - layer.3.k_cache 0.00131777 0.75816511 - layer.3.v_cache 0.00000216 0.00277180 - layer.4.k_cache 0.00349415 1.36191077 - layer.4.v_cache 0.00000336 0.00467361 - layer.4.output 0.00015143 0.09590540 - ------------------------------------------------------------------------------------- - TOTAL 0.00240374 1.69853098 - (elements=4,358,144) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4358144 -Total Bytes 514404 -BPFP 0.9443 bits/point -EBPFP 1.8885 equivalent bits/point -MSE 1.698531 ----------------------- -------------------------------------------------------- -Time: 0.937s Load: 0.016s, Pack+Encode: 0.348s, Decode+Unpack: 0.573s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 304, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 304, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6985 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample79-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample79-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample8-layer4-item1.zst (89/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample8-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 351, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 351, 128) -Output shape: (1, 351, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) -> torch.Size([1, 1, 351, 1024]) - layer.4.output: torch.Size([1, 351, 4096]) -> torch.Size([1, 1, 351, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,312B, BPFP=0.2073 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,888B, BPFP=1.3107 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,196B, BPFP=0.6276 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,544B, BPFP=1.3921 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,020B, BPFP=0.8017 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,688B, BPFP=1.4176 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,084B, BPFP=0.8699 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,024B, BPFP=1.3805 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,436B, BPFP=0.6107 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,476B, BPFP=1.4351 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,856B, BPFP=0.8060 -⌛️ [2/4] FRONTEND: Frontend time: 0.401s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.output: torch.Size([1, 351, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 351, 128]) - layer.0.v_cache: torch.Size([1, 8, 351, 128]) - layer.1.k_cache: torch.Size([1, 8, 351, 128]) - layer.1.v_cache: torch.Size([1, 8, 351, 128]) - layer.2.k_cache: torch.Size([1, 8, 351, 128]) - layer.2.v_cache: torch.Size([1, 8, 351, 128]) - layer.3.k_cache: torch.Size([1, 8, 351, 128]) - layer.3.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.k_cache: torch.Size([1, 8, 351, 128]) - layer.4.v_cache: torch.Size([1, 8, 351, 128]) - layer.4.output: torch.Size([1, 351, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02287757 17.95461488 - layer.0.v_cache 0.00000026 0.00029261 - layer.1.k_cache 0.00288319 1.24003249 - layer.1.v_cache 0.00000079 0.00108927 - layer.2.k_cache 0.00116994 0.69094001 - layer.2.v_cache 0.00000119 0.00168178 - layer.3.k_cache 0.00131288 0.75587329 - layer.3.v_cache 0.00000224 0.00277476 - layer.4.k_cache 0.00358885 1.38782269 - layer.4.v_cache 0.00000339 0.00464610 - layer.4.output 0.00014945 0.08938548 - ------------------------------------------------------------------------------------- - TOTAL 0.00231701 1.59980784 - (elements=5,031,936) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5031936 -Total Bytes 596524 -BPFP 0.9484 bits/point -EBPFP 1.8968 equivalent bits/point -MSE 1.599808 ----------------------- -------------------------------------------------------- -Time: 1.097s Load: 0.018s, Pack+Encode: 0.401s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 351, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 351, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5998 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample8-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample8-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample80-layer4-item1.zst (90/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample80-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 333, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 333, 128) -Output shape: (1, 333, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) -> torch.Size([1, 1, 333, 1024]) - layer.4.output: torch.Size([1, 333, 4096]) -> torch.Size([1, 1, 333, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,020B, BPFP=0.2116 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,816B, BPFP=1.3330 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,396B, BPFP=0.6427 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,040B, BPFP=1.4555 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,528B, BPFP=0.8335 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,588B, BPFP=1.4918 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,124B, BPFP=0.8944 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,024B, BPFP=1.4551 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,120B, BPFP=0.6363 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,248B, BPFP=1.5073 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 143,320B, BPFP=0.8406 -⌛️ [2/4] FRONTEND: Frontend time: 0.404s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.output: torch.Size([1, 333, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.677s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 333, 128]) - layer.0.v_cache: torch.Size([1, 8, 333, 128]) - layer.1.k_cache: torch.Size([1, 8, 333, 128]) - layer.1.v_cache: torch.Size([1, 8, 333, 128]) - layer.2.k_cache: torch.Size([1, 8, 333, 128]) - layer.2.v_cache: torch.Size([1, 8, 333, 128]) - layer.3.k_cache: torch.Size([1, 8, 333, 128]) - layer.3.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.k_cache: torch.Size([1, 8, 333, 128]) - layer.4.v_cache: torch.Size([1, 8, 333, 128]) - layer.4.output: torch.Size([1, 333, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02328842 19.88035350 - layer.0.v_cache 0.00000026 0.00030554 - layer.1.k_cache 0.00290993 1.25292795 - layer.1.v_cache 0.00000078 0.00112473 - layer.2.k_cache 0.00115179 0.70530733 - layer.2.v_cache 0.00000114 0.00169615 - layer.3.k_cache 0.00134347 0.76629785 - layer.3.v_cache 0.00000212 0.00275283 - layer.4.k_cache 0.00362086 1.42940157 - layer.4.v_cache 0.00000365 0.00458617 - layer.4.output 0.00013924 0.09313003 - ------------------------------------------------------------------------------------- - TOTAL 0.00234853 1.74409098 - (elements=4,773,888) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4773888 -Total Bytes 589224 -BPFP 0.9874 bits/point -EBPFP 1.9748 equivalent bits/point -MSE 1.744091 ----------------------- -------------------------------------------------------- -Time: 1.099s Load: 0.017s, Pack+Encode: 0.404s, Decode+Unpack: 0.677s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 333, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 333, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7441 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample80-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample80-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample81-layer4-item1.zst (91/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample81-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 319, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 319, 128) -Output shape: (1, 319, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.0.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.1.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.1.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.2.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.2.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.3.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.3.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.4.k_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.4.v_cache: torch.Size([1, 8, 319, 128]) -> torch.Size([1, 1, 319, 1024]) - layer.4.output: torch.Size([1, 319, 4096]) -> torch.Size([1, 1, 319, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 7,920B, BPFP=0.1940 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 50,024B, BPFP=1.2251 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,908B, BPFP=0.5855 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,956B, BPFP=1.3214 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,176B, BPFP=0.7880 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 55,376B, BPFP=1.3562 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,868B, BPFP=0.8294 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 53,824B, BPFP=1.3182 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,764B, BPFP=0.5820 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 56,032B, BPFP=1.3723 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 125,548B, BPFP=0.7687 -⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 319, 128]) - layer.0.v_cache: torch.Size([1, 8, 319, 128]) - layer.1.k_cache: torch.Size([1, 8, 319, 128]) - layer.1.v_cache: torch.Size([1, 8, 319, 128]) - layer.2.k_cache: torch.Size([1, 8, 319, 128]) - layer.2.v_cache: torch.Size([1, 8, 319, 128]) - layer.3.k_cache: torch.Size([1, 8, 319, 128]) - layer.3.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.k_cache: torch.Size([1, 8, 319, 128]) - layer.4.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.output: torch.Size([1, 319, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.579s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 319, 128]) - layer.0.v_cache: torch.Size([1, 8, 319, 128]) - layer.1.k_cache: torch.Size([1, 8, 319, 128]) - layer.1.v_cache: torch.Size([1, 8, 319, 128]) - layer.2.k_cache: torch.Size([1, 8, 319, 128]) - layer.2.v_cache: torch.Size([1, 8, 319, 128]) - layer.3.k_cache: torch.Size([1, 8, 319, 128]) - layer.3.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.k_cache: torch.Size([1, 8, 319, 128]) - layer.4.v_cache: torch.Size([1, 8, 319, 128]) - layer.4.output: torch.Size([1, 319, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02501136 17.45212089 - layer.0.v_cache 0.00000027 0.00029813 - layer.1.k_cache 0.00284013 1.11586721 - layer.1.v_cache 0.00000084 0.00114060 - layer.2.k_cache 0.00117936 0.68014794 - layer.2.v_cache 0.00000120 0.00174344 - layer.3.k_cache 0.00130692 0.74628202 - layer.3.v_cache 0.00000226 0.00279442 - layer.4.k_cache 0.00361670 1.37434966 - layer.4.v_cache 0.00000371 0.00476775 - layer.4.output 0.00016871 0.09387816 - ------------------------------------------------------------------------------------- - TOTAL 0.00247411 1.55393034 - (elements=4,573,184) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4573184 -Total Bytes 516396 -BPFP 0.9033 bits/point -EBPFP 1.8067 equivalent bits/point -MSE 1.553930 ----------------------- -------------------------------------------------------- -Time: 0.955s Load: 0.018s, Pack+Encode: 0.358s, Decode+Unpack: 0.579s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 319, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 319, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5539 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample81-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample81-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample82-layer4-item1.zst (92/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample82-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 353, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 353, 128) -Output shape: (1, 353, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.0.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.1.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.1.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.2.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.2.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.3.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.3.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.4.k_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.4.v_cache: torch.Size([1, 8, 353, 128]) -> torch.Size([1, 1, 353, 1024]) - layer.4.output: torch.Size([1, 353, 4096]) -> torch.Size([1, 1, 353, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,340B, BPFP=0.2067 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 59,260B, BPFP=1.3115 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,892B, BPFP=0.6173 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,500B, BPFP=1.3832 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,356B, BPFP=0.8046 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,312B, BPFP=1.4012 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,152B, BPFP=0.8665 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,580B, BPFP=1.3629 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,928B, BPFP=0.5960 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,124B, BPFP=1.4192 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 149,720B, BPFP=0.8284 -⌛️ [2/4] FRONTEND: Frontend time: 0.403s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 353, 128]) - layer.0.v_cache: torch.Size([1, 8, 353, 128]) - layer.1.k_cache: torch.Size([1, 8, 353, 128]) - layer.1.v_cache: torch.Size([1, 8, 353, 128]) - layer.2.k_cache: torch.Size([1, 8, 353, 128]) - layer.2.v_cache: torch.Size([1, 8, 353, 128]) - layer.3.k_cache: torch.Size([1, 8, 353, 128]) - layer.3.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.k_cache: torch.Size([1, 8, 353, 128]) - layer.4.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.output: torch.Size([1, 353, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.683s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 353, 128]) - layer.0.v_cache: torch.Size([1, 8, 353, 128]) - layer.1.k_cache: torch.Size([1, 8, 353, 128]) - layer.1.v_cache: torch.Size([1, 8, 353, 128]) - layer.2.k_cache: torch.Size([1, 8, 353, 128]) - layer.2.v_cache: torch.Size([1, 8, 353, 128]) - layer.3.k_cache: torch.Size([1, 8, 353, 128]) - layer.3.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.k_cache: torch.Size([1, 8, 353, 128]) - layer.4.v_cache: torch.Size([1, 8, 353, 128]) - layer.4.output: torch.Size([1, 353, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02381471 17.66025031 - layer.0.v_cache 0.00000027 0.00031029 - layer.1.k_cache 0.00292436 1.18357983 - layer.1.v_cache 0.00000085 0.00116068 - layer.2.k_cache 0.00118067 0.69342292 - layer.2.v_cache 0.00000120 0.00168787 - layer.3.k_cache 0.00130453 0.76740297 - layer.3.v_cache 0.00000223 0.00282972 - layer.4.k_cache 0.00350570 1.37874994 - layer.4.v_cache 0.00000328 0.00468553 - layer.4.output 0.00016368 0.10244774 - ------------------------------------------------------------------------------------- - TOTAL 0.00238518 1.57884793 - (elements=5,060,608) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5060608 -Total Bytes 600164 -BPFP 0.9488 bits/point -EBPFP 1.8975 equivalent bits/point -MSE 1.578848 ----------------------- -------------------------------------------------------- -Time: 1.103s Load: 0.017s, Pack+Encode: 0.403s, Decode+Unpack: 0.683s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 353, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 353, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5788 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample82-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample82-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample83-layer4-item1.zst (93/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample83-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.020s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 344, 128) -Output shape: (1, 344, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) -> torch.Size([1, 1, 344, 1024]) - layer.4.output: torch.Size([1, 344, 4096]) -> torch.Size([1, 1, 344, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,196B, BPFP=0.2088 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 60,076B, BPFP=1.3644 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,624B, BPFP=0.6274 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 64,208B, BPFP=1.4582 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,768B, BPFP=0.8350 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 65,428B, BPFP=1.4859 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,628B, BPFP=0.9000 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,880B, BPFP=1.4508 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,396B, BPFP=0.6222 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 66,320B, BPFP=1.5062 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 151,360B, BPFP=0.8594 -⌛️ [2/4] FRONTEND: Frontend time: 0.408s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 344, 128]) - layer.0.v_cache: torch.Size([1, 8, 344, 128]) - layer.1.k_cache: torch.Size([1, 8, 344, 128]) - layer.1.v_cache: torch.Size([1, 8, 344, 128]) - layer.2.k_cache: torch.Size([1, 8, 344, 128]) - layer.2.v_cache: torch.Size([1, 8, 344, 128]) - layer.3.k_cache: torch.Size([1, 8, 344, 128]) - layer.3.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.k_cache: torch.Size([1, 8, 344, 128]) - layer.4.v_cache: torch.Size([1, 8, 344, 128]) - layer.4.output: torch.Size([1, 344, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02504009 19.21394951 - layer.0.v_cache 0.00000027 0.00032140 - layer.1.k_cache 0.00288154 1.30633226 - layer.1.v_cache 0.00000082 0.00118759 - layer.2.k_cache 0.00115445 0.69845058 - layer.2.v_cache 0.00000117 0.00171946 - layer.3.k_cache 0.00131257 0.77306925 - layer.3.v_cache 0.00000224 0.00288045 - layer.4.k_cache 0.00356118 1.36480748 - layer.4.v_cache 0.00000339 0.00471342 - layer.4.output 0.00014496 0.09414457 - ------------------------------------------------------------------------------------- - TOTAL 0.00246697 1.69600069 - (elements=4,931,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4931584 -Total Bytes 611884 -BPFP 0.9926 bits/point -EBPFP 1.9852 equivalent bits/point -MSE 1.696001 ----------------------- -------------------------------------------------------- -Time: 1.106s Load: 0.020s, Pack+Encode: 0.408s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 344, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 344, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6960 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample83-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample83-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample85-layer4-item1.zst (94/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample85-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 321, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 321, 128) -Output shape: (1, 321, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.0.v_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.1.k_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.1.v_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.2.k_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.2.v_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.3.k_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.3.v_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.4.k_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.4.v_cache: torch.Size([1, 8, 321, 128]) -> torch.Size([1, 1, 321, 1024]) - layer.4.output: torch.Size([1, 321, 4096]) -> torch.Size([1, 1, 321, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,528B, BPFP=0.2076 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 55,100B, BPFP=1.3410 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,964B, BPFP=0.6562 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 59,644B, BPFP=1.4516 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,844B, BPFP=0.8480 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 60,524B, BPFP=1.4730 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,052B, BPFP=0.9018 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 59,324B, BPFP=1.4438 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,716B, BPFP=0.6502 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 61,708B, BPFP=1.5018 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,028B, BPFP=0.8824 -⌛️ [2/4] FRONTEND: Frontend time: 0.425s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 321, 128]) - layer.0.v_cache: torch.Size([1, 8, 321, 128]) - layer.1.k_cache: torch.Size([1, 8, 321, 128]) - layer.1.v_cache: torch.Size([1, 8, 321, 128]) - layer.2.k_cache: torch.Size([1, 8, 321, 128]) - layer.2.v_cache: torch.Size([1, 8, 321, 128]) - layer.3.k_cache: torch.Size([1, 8, 321, 128]) - layer.3.v_cache: torch.Size([1, 8, 321, 128]) - layer.4.k_cache: torch.Size([1, 8, 321, 128]) - layer.4.v_cache: torch.Size([1, 8, 321, 128]) - layer.4.output: torch.Size([1, 321, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.689s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 321, 128]) - layer.0.v_cache: torch.Size([1, 8, 321, 128]) - layer.1.k_cache: torch.Size([1, 8, 321, 128]) - layer.1.v_cache: torch.Size([1, 8, 321, 128]) - layer.2.k_cache: torch.Size([1, 8, 321, 128]) - layer.2.v_cache: torch.Size([1, 8, 321, 128]) - layer.3.k_cache: torch.Size([1, 8, 321, 128]) - layer.3.v_cache: torch.Size([1, 8, 321, 128]) - layer.4.k_cache: torch.Size([1, 8, 321, 128]) - layer.4.v_cache: torch.Size([1, 8, 321, 128]) - layer.4.output: torch.Size([1, 321, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02357271 18.83268533 - layer.0.v_cache 0.00000026 0.00029514 - layer.1.k_cache 0.00287206 1.17502525 - layer.1.v_cache 0.00000077 0.00110052 - layer.2.k_cache 0.00116438 0.70194931 - layer.2.v_cache 0.00000114 0.00166172 - layer.3.k_cache 0.00132116 0.74770781 - layer.3.v_cache 0.00000214 0.00269032 - layer.4.k_cache 0.00354096 1.39350639 - layer.4.v_cache 0.00000337 0.00464604 - layer.4.output 0.00013764 0.08794310 - ------------------------------------------------------------------------------------- - TOTAL 0.00235925 1.65807430 - (elements=4,601,856) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4601856 -Total Bytes 575432 -BPFP 1.0003 bits/point -EBPFP 2.0007 equivalent bits/point -MSE 1.658074 ----------------------- -------------------------------------------------------- -Time: 1.131s Load: 0.017s, Pack+Encode: 0.425s, Decode+Unpack: 0.689s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 321, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 321, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6581 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample85-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample85-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample86-layer4-item1.zst (95/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample86-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 329, 128) -Output shape: (1, 329, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.output: torch.Size([1, 329, 4096]) -> torch.Size([1, 1, 329, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,836B, BPFP=0.2098 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 56,964B, BPFP=1.3527 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,712B, BPFP=0.6581 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,796B, BPFP=1.4674 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,916B, BPFP=0.8529 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,440B, BPFP=1.5065 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,104B, BPFP=0.9048 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,060B, BPFP=1.4737 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,808B, BPFP=0.6603 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,976B, BPFP=1.5192 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 148,180B, BPFP=0.8797 -⌛️ [2/4] FRONTEND: Frontend time: 0.414s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.682s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02395787 19.94184555 - layer.0.v_cache 0.00000026 0.00030125 - layer.1.k_cache 0.00285945 1.25454239 - layer.1.v_cache 0.00000078 0.00114939 - layer.2.k_cache 0.00114467 0.69678017 - layer.2.v_cache 0.00000113 0.00168984 - layer.3.k_cache 0.00130198 0.76811756 - layer.3.v_cache 0.00000219 0.00277922 - layer.4.k_cache 0.00350851 1.41159679 - layer.4.v_cache 0.00000338 0.00477918 - layer.4.output 0.00014624 0.09134319 - ------------------------------------------------------------------------------------- - TOTAL 0.00238323 1.74635387 - (elements=4,716,544) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4716544 -Total Bytes 594792 -BPFP 1.0089 bits/point -EBPFP 2.0177 equivalent bits/point -MSE 1.746354 ----------------------- -------------------------------------------------------- -Time: 1.114s Load: 0.019s, Pack+Encode: 0.414s, Decode+Unpack: 0.682s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7464 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample86-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample86-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample87-layer4-item1.zst (96/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample87-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 350, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.019s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 350, 128) -Output shape: (1, 350, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.0.v_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.1.k_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.1.v_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.2.k_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.2.v_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.3.k_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.3.v_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.4.k_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.4.v_cache: torch.Size([1, 8, 350, 128]) -> torch.Size([1, 1, 350, 1024]) - layer.4.output: torch.Size([1, 350, 4096]) -> torch.Size([1, 1, 350, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,972B, BPFP=0.2003 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,568B, BPFP=1.2850 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,916B, BPFP=0.6231 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,136B, BPFP=1.3870 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,932B, BPFP=0.8021 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 62,520B, BPFP=1.3955 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,312B, BPFP=0.8775 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,444B, BPFP=1.3715 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,060B, BPFP=0.6040 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,772B, BPFP=1.4235 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 136,736B, BPFP=0.7630 -⌛️ [2/4] FRONTEND: Frontend time: 0.404s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 350, 128]) - layer.0.v_cache: torch.Size([1, 8, 350, 128]) - layer.1.k_cache: torch.Size([1, 8, 350, 128]) - layer.1.v_cache: torch.Size([1, 8, 350, 128]) - layer.2.k_cache: torch.Size([1, 8, 350, 128]) - layer.2.v_cache: torch.Size([1, 8, 350, 128]) - layer.3.k_cache: torch.Size([1, 8, 350, 128]) - layer.3.v_cache: torch.Size([1, 8, 350, 128]) - layer.4.k_cache: torch.Size([1, 8, 350, 128]) - layer.4.v_cache: torch.Size([1, 8, 350, 128]) - layer.4.output: torch.Size([1, 350, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.678s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 350, 128]) - layer.0.v_cache: torch.Size([1, 8, 350, 128]) - layer.1.k_cache: torch.Size([1, 8, 350, 128]) - layer.1.v_cache: torch.Size([1, 8, 350, 128]) - layer.2.k_cache: torch.Size([1, 8, 350, 128]) - layer.2.v_cache: torch.Size([1, 8, 350, 128]) - layer.3.k_cache: torch.Size([1, 8, 350, 128]) - layer.3.v_cache: torch.Size([1, 8, 350, 128]) - layer.4.k_cache: torch.Size([1, 8, 350, 128]) - layer.4.v_cache: torch.Size([1, 8, 350, 128]) - layer.4.output: torch.Size([1, 350, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02252116 17.66507673 - layer.0.v_cache 0.00000028 0.00030418 - layer.1.k_cache 0.00283737 1.20403015 - layer.1.v_cache 0.00000078 0.00103736 - layer.2.k_cache 0.00114986 0.68491377 - layer.2.v_cache 0.00000111 0.00154586 - layer.3.k_cache 0.00128401 0.75009391 - layer.3.v_cache 0.00000216 0.00257724 - layer.4.k_cache 0.00352053 1.36879586 - layer.4.v_cache 0.00000313 0.00438053 - layer.4.output 0.00014459 0.09284332 - ------------------------------------------------------------------------------------- - TOTAL 0.00227848 1.57529492 - (elements=5,017,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 5017600 -Total Bytes 583368 -BPFP 0.9301 bits/point -EBPFP 1.8602 equivalent bits/point -MSE 1.575295 ----------------------- -------------------------------------------------------- -Time: 1.101s Load: 0.019s, Pack+Encode: 0.404s, Decode+Unpack: 0.678s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 350, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 350, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5753 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample87-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample87-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample89-layer4-item1.zst (97/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample89-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 323, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.018s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 323, 128) -Output shape: (1, 323, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) -> torch.Size([1, 1, 323, 1024]) - layer.4.output: torch.Size([1, 323, 4096]) -> torch.Size([1, 1, 323, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,880B, BPFP=0.2148 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 55,280B, BPFP=1.3371 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,460B, BPFP=0.6642 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 59,856B, BPFP=1.4478 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 35,340B, BPFP=0.8548 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 61,108B, BPFP=1.4780 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,672B, BPFP=0.9112 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 59,748B, BPFP=1.4451 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,340B, BPFP=0.6371 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 61,632B, BPFP=1.4907 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 146,224B, BPFP=0.8842 -⌛️ [2/4] FRONTEND: Frontend time: 0.406s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.output: torch.Size([1, 323, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.683s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 323, 128]) - layer.0.v_cache: torch.Size([1, 8, 323, 128]) - layer.1.k_cache: torch.Size([1, 8, 323, 128]) - layer.1.v_cache: torch.Size([1, 8, 323, 128]) - layer.2.k_cache: torch.Size([1, 8, 323, 128]) - layer.2.v_cache: torch.Size([1, 8, 323, 128]) - layer.3.k_cache: torch.Size([1, 8, 323, 128]) - layer.3.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.k_cache: torch.Size([1, 8, 323, 128]) - layer.4.v_cache: torch.Size([1, 8, 323, 128]) - layer.4.output: torch.Size([1, 323, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02344667 19.80046228 - layer.0.v_cache 0.00000026 0.00030191 - layer.1.k_cache 0.00293942 1.18681926 - layer.1.v_cache 0.00000088 0.00114852 - layer.2.k_cache 0.00118806 0.70164608 - layer.2.v_cache 0.00000115 0.00169712 - layer.3.k_cache 0.00133183 0.77378250 - layer.3.v_cache 0.00000219 0.00281984 - layer.4.k_cache 0.00365354 1.41532430 - layer.4.v_cache 0.00000330 0.00467146 - layer.4.output 0.00014070 0.09392002 - ------------------------------------------------------------------------------------- - TOTAL 0.00236644 1.73316810 - (elements=4,630,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4630528 -Total Bytes 579540 -BPFP 1.0013 bits/point -EBPFP 2.0025 equivalent bits/point -MSE 1.733168 ----------------------- -------------------------------------------------------- -Time: 1.107s Load: 0.018s, Pack+Encode: 0.406s, Decode+Unpack: 0.683s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 323, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 323, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7332 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample89-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample89-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample9-layer4-item1.zst (98/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample9-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 419, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.022s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 419, 128) -Output shape: (1, 419, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.0.v_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.1.k_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.1.v_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.2.k_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.2.v_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.3.k_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.3.v_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.4.k_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.4.v_cache: torch.Size([1, 8, 419, 128]) -> torch.Size([1, 1, 419, 1024]) - layer.4.output: torch.Size([1, 419, 4096]) -> torch.Size([1, 1, 419, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 10,296B, BPFP=0.1920 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,564B, BPFP=1.2784 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 33,004B, BPFP=0.6154 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,156B, BPFP=1.3640 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 42,316B, BPFP=0.7890 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,392B, BPFP=1.3684 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 46,156B, BPFP=0.8606 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,396B, BPFP=1.3499 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 32,080B, BPFP=0.5982 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,200B, BPFP=1.4021 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 157,804B, BPFP=0.7356 -⌛️ [2/4] FRONTEND: Frontend time: 0.464s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 419, 128]) - layer.0.v_cache: torch.Size([1, 8, 419, 128]) - layer.1.k_cache: torch.Size([1, 8, 419, 128]) - layer.1.v_cache: torch.Size([1, 8, 419, 128]) - layer.2.k_cache: torch.Size([1, 8, 419, 128]) - layer.2.v_cache: torch.Size([1, 8, 419, 128]) - layer.3.k_cache: torch.Size([1, 8, 419, 128]) - layer.3.v_cache: torch.Size([1, 8, 419, 128]) - layer.4.k_cache: torch.Size([1, 8, 419, 128]) - layer.4.v_cache: torch.Size([1, 8, 419, 128]) - layer.4.output: torch.Size([1, 419, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.763s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 419, 128]) - layer.0.v_cache: torch.Size([1, 8, 419, 128]) - layer.1.k_cache: torch.Size([1, 8, 419, 128]) - layer.1.v_cache: torch.Size([1, 8, 419, 128]) - layer.2.k_cache: torch.Size([1, 8, 419, 128]) - layer.2.v_cache: torch.Size([1, 8, 419, 128]) - layer.3.k_cache: torch.Size([1, 8, 419, 128]) - layer.3.v_cache: torch.Size([1, 8, 419, 128]) - layer.4.k_cache: torch.Size([1, 8, 419, 128]) - layer.4.v_cache: torch.Size([1, 8, 419, 128]) - layer.4.output: torch.Size([1, 419, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02248870 17.83592818 - layer.0.v_cache 0.00000028 0.00030130 - layer.1.k_cache 0.00282601 1.16490869 - layer.1.v_cache 0.00000075 0.00101974 - layer.2.k_cache 0.00117249 0.67280830 - layer.2.v_cache 0.00000110 0.00148266 - layer.3.k_cache 0.00129145 0.74123038 - layer.3.v_cache 0.00000206 0.00253224 - layer.4.k_cache 0.00360397 1.36288867 - layer.4.v_cache 0.00000307 0.00431818 - layer.4.output 0.00013808 0.08354907 - ------------------------------------------------------------------------------------- - TOTAL 0.00228159 1.58011533 - (elements=6,006,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 6006784 -Total Bytes 684364 -BPFP 0.9115 bits/point -EBPFP 1.8229 equivalent bits/point -MSE 1.580115 ----------------------- -------------------------------------------------------- -Time: 1.249s Load: 0.022s, Pack+Encode: 0.464s, Decode+Unpack: 0.763s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 419, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 419, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.5801 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample9-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample9-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample93-layer4-item1.zst (99/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample93-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 336, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 336, 128) -Output shape: (1, 336, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) -> torch.Size([1, 1, 336, 1024]) - layer.4.output: torch.Size([1, 336, 4096]) -> torch.Size([1, 1, 336, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 9,084B, BPFP=0.2112 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,116B, BPFP=1.3513 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,116B, BPFP=0.6537 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 62,396B, BPFP=1.4508 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,184B, BPFP=0.8413 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,956B, BPFP=1.4871 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 39,024B, BPFP=0.9074 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,448B, BPFP=1.4520 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,936B, BPFP=0.6496 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 64,872B, BPFP=1.5084 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 144,528B, BPFP=0.8401 -⌛️ [2/4] FRONTEND: Frontend time: 0.423s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.output: torch.Size([1, 336, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.682s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 336, 128]) - layer.0.v_cache: torch.Size([1, 8, 336, 128]) - layer.1.k_cache: torch.Size([1, 8, 336, 128]) - layer.1.v_cache: torch.Size([1, 8, 336, 128]) - layer.2.k_cache: torch.Size([1, 8, 336, 128]) - layer.2.v_cache: torch.Size([1, 8, 336, 128]) - layer.3.k_cache: torch.Size([1, 8, 336, 128]) - layer.3.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.k_cache: torch.Size([1, 8, 336, 128]) - layer.4.v_cache: torch.Size([1, 8, 336, 128]) - layer.4.output: torch.Size([1, 336, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02369028 18.95129249 - layer.0.v_cache 0.00000026 0.00030451 - layer.1.k_cache 0.00291391 1.25825728 - layer.1.v_cache 0.00000078 0.00114665 - layer.2.k_cache 0.00115383 0.69948632 - layer.2.v_cache 0.00000117 0.00169915 - layer.3.k_cache 0.00133386 0.77325049 - layer.3.v_cache 0.00000214 0.00277408 - layer.4.k_cache 0.00353874 1.42218499 - layer.4.v_cache 0.00000342 0.00473520 - layer.4.output 0.00014590 0.09810351 - ------------------------------------------------------------------------------------- - TOTAL 0.00237300 1.67911037 - (elements=4,816,896) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4816896 -Total Bytes 596660 -BPFP 0.9909 bits/point -EBPFP 1.9819 equivalent bits/point -MSE 1.679110 ----------------------- -------------------------------------------------------- -Time: 1.122s Load: 0.017s, Pack+Encode: 0.423s, Decode+Unpack: 0.682s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 336, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 336, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.6791 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample93-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample93-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample97-layer4-item1.zst (100/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample97-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.017s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 329, 128) -Output shape: (1, 329, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) -> torch.Size([1, 1, 329, 1024]) - layer.4.output: torch.Size([1, 329, 4096]) -> torch.Size([1, 1, 329, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 8,840B, BPFP=0.2099 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,008B, BPFP=1.3537 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,508B, BPFP=0.6532 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,820B, BPFP=1.4680 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,144B, BPFP=0.8583 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,364B, BPFP=1.5047 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 38,052B, BPFP=0.9036 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,592B, BPFP=1.4626 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,744B, BPFP=0.6588 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,972B, BPFP=1.5191 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 137,280B, BPFP=0.8150 -⌛️ [2/4] FRONTEND: Frontend time: 0.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) -⌛️ [3/4] BACKEND: Backend time: 0.680s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 329, 128]) - layer.0.v_cache: torch.Size([1, 8, 329, 128]) - layer.1.k_cache: torch.Size([1, 8, 329, 128]) - layer.1.v_cache: torch.Size([1, 8, 329, 128]) - layer.2.k_cache: torch.Size([1, 8, 329, 128]) - layer.2.v_cache: torch.Size([1, 8, 329, 128]) - layer.3.k_cache: torch.Size([1, 8, 329, 128]) - layer.3.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.k_cache: torch.Size([1, 8, 329, 128]) - layer.4.v_cache: torch.Size([1, 8, 329, 128]) - layer.4.output: torch.Size([1, 329, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02541956 20.29252499 - layer.0.v_cache 0.00000027 0.00030812 - layer.1.k_cache 0.00292153 1.22033998 - layer.1.v_cache 0.00000078 0.00116398 - layer.2.k_cache 0.00117184 0.70228094 - layer.2.v_cache 0.00000114 0.00171019 - layer.3.k_cache 0.00131754 0.75974851 - layer.3.v_cache 0.00000217 0.00282986 - layer.4.k_cache 0.00353394 1.40795481 - layer.4.v_cache 0.00000342 0.00474274 - layer.4.output 0.00014204 0.09298673 - ------------------------------------------------------------------------------------- - TOTAL 0.00249574 1.76896793 - (elements=4,716,544) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture hyperprior-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4716544 -Total Bytes 583324 -BPFP 0.9894 bits/point -EBPFP 1.9788 equivalent bits/point -MSE 1.768968 ----------------------- -------------------------------------------------------- -Time: 1.104s Load: 0.017s, Pack+Encode: 0.407s, Decode+Unpack: 0.680s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 329, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 329, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.7690 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_hellaswag/sample97-layer4-item1.zst - to output-fixed/qwen/lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample97-layer4-item1.zst ------------------------- ---------------------------- -TOTAL PROCESSING SUMMARY ------------------------- ---------------------------- -Total files 100 -Avg BPFP 0.9682 bits/point -Avg EBPFP 1.9364 equivalent bits/point -Avg MSE 1.674582 -Avg Time 1.092s ------------------------- ---------------------------- +version https://git-lfs.github.com/spec/v1 +oid sha256:8639a396373412beaf30ae5de6f401b5d22852388c70f8ca4237f3e5c309c00e +size 1124952