diff --git a/.gitattributes b/.gitattributes index 608aa0a2254d0f8174133cd8e49329dbfa532a58..32b53dc46d928cb4b6125e91e8b7e3e8c2812027 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5133,3 +5133,7 @@ lambda0.02/elic-featurecoding-8bit-individual/fc_arc_challenge/sample97-layer4-i lambda0.02/elic-featurecoding-8bit-individual/fc_gsm8k/sample969-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text lambda0.02/elic-featurecoding-8bit-individual/fc_hellaswag/sample10-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text lambda0.02/elic-featurecoding-8bit-individual/fc_hellaswag/sample132-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample19-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample33-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample45-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample99-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text diff --git a/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log index 8c85e42a54b347cd5cc0e0d71e3e84ff2b2da878..2ac53441a867d8f10d179caa313e150989aa5100 100644 --- a/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log +++ b/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log @@ -1,16958 +1,3 @@ -Experiment: dtufc_elic-featurecoding_qwen_individual -Log file: output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log -DTUFCCodecConfig: - arch: elic-featurecoding - handler: qwen - checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_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/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar -Checkpoint epoch: 520 -Loaded elic-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 elic-featurecoding -Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_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_truthfulqa_mc1 -Output output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1 ----------------- ------------------------------------------------------------------------------------------------------------------- -Files found: 100 ----------------------------------------------------------------------- - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample1-layer4-item1.zst (1/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample1-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 276, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 276, 128) -Output shape: (1, 276, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.0.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.1.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.1.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.2.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.2.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.3.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.3.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.4.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.4.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.4.output: torch.Size([1, 276, 4096]) -> torch.Size([1, 1, 276, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 36,608B, BPFP=1.0362 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 108,544B, BPFP=3.0725 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 69,812B, BPFP=1.9761 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 116,108B, BPFP=3.2866 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 83,476B, BPFP=2.3629 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 117,560B, BPFP=3.3277 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 84,516B, BPFP=2.3923 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 113,792B, BPFP=3.2210 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 69,508B, BPFP=1.9675 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 117,660B, BPFP=3.3305 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 293,832B, BPFP=2.0793 -⌛️ [2/4] FRONTEND: Frontend time: 2.780s (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, 276, 128]) - layer.0.v_cache: torch.Size([1, 8, 276, 128]) - layer.1.k_cache: torch.Size([1, 8, 276, 128]) - layer.1.v_cache: torch.Size([1, 8, 276, 128]) - layer.2.k_cache: torch.Size([1, 8, 276, 128]) - layer.2.v_cache: torch.Size([1, 8, 276, 128]) - layer.3.k_cache: torch.Size([1, 8, 276, 128]) - layer.3.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.k_cache: torch.Size([1, 8, 276, 128]) - layer.4.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.output: torch.Size([1, 276, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.793s - -[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, 276, 128]) - layer.0.v_cache: torch.Size([1, 8, 276, 128]) - layer.1.k_cache: torch.Size([1, 8, 276, 128]) - layer.1.v_cache: torch.Size([1, 8, 276, 128]) - layer.2.k_cache: torch.Size([1, 8, 276, 128]) - layer.2.v_cache: torch.Size([1, 8, 276, 128]) - layer.3.k_cache: torch.Size([1, 8, 276, 128]) - layer.3.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.k_cache: torch.Size([1, 8, 276, 128]) - layer.4.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.output: torch.Size([1, 276, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02420060 4.27292312 - layer.0.v_cache 0.00000027 0.00005560 - layer.1.k_cache 0.00304789 0.35549645 - layer.1.v_cache 0.00000080 0.00017567 - layer.2.k_cache 0.00119192 0.15572262 - layer.2.v_cache 0.00000117 0.00024789 - layer.3.k_cache 0.00127914 0.16926714 - layer.3.v_cache 0.00000214 0.00040433 - layer.4.k_cache 0.00374547 0.40130817 - layer.4.v_cache 0.00000302 0.00064129 - layer.4.output 0.00013416 0.02505591 - ------------------------------------------------------------------------------------- - TOTAL 0.00242922 0.38974757 - (elements=3,956,736) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3956736 -Total Bytes 1211416 -BPFP 2.4493 bits/point -EBPFP 4.8986 equivalent bits/point -MSE 0.389748 ----------------------- -------------------------------------------------------- -Time: 4.586s Load: 0.013s, Pack+Encode: 2.780s, Decode+Unpack: 1.793s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 276, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3897 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample1-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample1-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample10-layer4-item1.zst (2/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample10-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 283, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 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, 283, 128) -Output shape: (1, 283, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.0.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.1.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.1.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.2.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.2.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.3.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.3.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.4.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.4.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.4.output: torch.Size([1, 283, 4096]) -> torch.Size([1, 1, 283, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 34,292B, BPFP=0.9467 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 108,740B, BPFP=3.0019 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 71,700B, BPFP=1.9794 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 115,352B, BPFP=3.1844 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 84,052B, BPFP=2.3203 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 118,080B, BPFP=3.2597 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 85,648B, BPFP=2.3644 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 115,816B, BPFP=3.1972 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 70,552B, BPFP=1.9477 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 118,688B, BPFP=3.2765 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 309,612B, BPFP=2.1368 -⌛️ [2/4] FRONTEND: Frontend time: 2.349s (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, 283, 128]) - layer.0.v_cache: torch.Size([1, 8, 283, 128]) - layer.1.k_cache: torch.Size([1, 8, 283, 128]) - layer.1.v_cache: torch.Size([1, 8, 283, 128]) - layer.2.k_cache: torch.Size([1, 8, 283, 128]) - layer.2.v_cache: torch.Size([1, 8, 283, 128]) - layer.3.k_cache: torch.Size([1, 8, 283, 128]) - layer.3.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.k_cache: torch.Size([1, 8, 283, 128]) - layer.4.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.output: torch.Size([1, 283, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.780s - -[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, 283, 128]) - layer.0.v_cache: torch.Size([1, 8, 283, 128]) - layer.1.k_cache: torch.Size([1, 8, 283, 128]) - layer.1.v_cache: torch.Size([1, 8, 283, 128]) - layer.2.k_cache: torch.Size([1, 8, 283, 128]) - layer.2.v_cache: torch.Size([1, 8, 283, 128]) - layer.3.k_cache: torch.Size([1, 8, 283, 128]) - layer.3.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.k_cache: torch.Size([1, 8, 283, 128]) - layer.4.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.output: torch.Size([1, 283, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02559670 4.31552631 - layer.0.v_cache 0.00000030 0.00005647 - layer.1.k_cache 0.00287003 0.33137545 - layer.1.v_cache 0.00000078 0.00017479 - layer.2.k_cache 0.00118918 0.15545842 - layer.2.v_cache 0.00000127 0.00025329 - layer.3.k_cache 0.00127329 0.16914411 - layer.3.v_cache 0.00000227 0.00041313 - layer.4.k_cache 0.00368869 0.43204250 - layer.4.v_cache 0.00000310 0.00063960 - layer.4.output 0.00017183 0.02778237 - ------------------------------------------------------------------------------------- - TOTAL 0.00252235 0.39401525 - (elements=4,057,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4057088 -Total Bytes 1232532 -BPFP 2.4304 bits/point -EBPFP 4.8608 equivalent bits/point -MSE 0.394015 ----------------------- -------------------------------------------------------- -Time: 4.143s Load: 0.015s, Pack+Encode: 2.349s, Decode+Unpack: 1.780s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 283, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3940 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample10-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample10-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample100-layer4-item1.zst (3/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample100-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 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, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 20,716B, BPFP=1.0243 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,200B, BPFP=3.2733 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,260B, BPFP=2.0402 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 69,496B, BPFP=3.4363 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,392B, BPFP=2.4917 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 70,608B, BPFP=3.4913 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,804B, BPFP=2.5121 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 69,208B, BPFP=3.4221 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,616B, BPFP=2.1072 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,092B, BPFP=3.5152 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 187,984B, BPFP=2.3238 -⌛️ [2/4] FRONTEND: Frontend time: 2.127s (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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.387s - -[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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02645765 4.01256204 - layer.0.v_cache 0.00000027 0.00005962 - layer.1.k_cache 0.00311677 0.32818024 - layer.1.v_cache 0.00000085 0.00018473 - layer.2.k_cache 0.00117315 0.16621186 - layer.2.v_cache 0.00000112 0.00026890 - layer.3.k_cache 0.00135825 0.18286553 - layer.3.v_cache 0.00000212 0.00042073 - layer.4.k_cache 0.00352137 0.46430095 - layer.4.v_cache 0.00000299 0.00067519 - layer.4.output 0.00021777 0.04170992 - ------------------------------------------------------------------------------------- - TOTAL 0.00260754 0.38018353 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 740376 -BPFP 2.6149 bits/point -EBPFP 5.2298 equivalent bits/point -MSE 0.380184 ----------------------- -------------------------------------------------------- -Time: 3.522s Load: 0.009s, Pack+Encode: 2.127s, Decode+Unpack: 1.387s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3802 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample100-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample100-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample101-layer4-item1.zst (4/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample101-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: 0.009s - ------------------------------------------------------------- -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: 20,900B, BPFP=1.0603 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,628B, BPFP=3.3801 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,336B, BPFP=2.0970 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 69,848B, BPFP=3.5434 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,280B, BPFP=2.5507 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,064B, BPFP=3.6051 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,724B, BPFP=2.5733 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,588B, BPFP=3.4795 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,116B, BPFP=2.1366 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,132B, BPFP=3.6086 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 189,500B, BPFP=2.4034 -⌛️ [2/4] FRONTEND: Frontend time: 1.891s (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: 1.380s - -[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 0.02866718 4.12000899 - layer.0.v_cache 0.00000026 0.00005842 - layer.1.k_cache 0.00317794 0.32632902 - layer.1.v_cache 0.00000104 0.00020412 - layer.2.k_cache 0.00114871 0.17073707 - layer.2.v_cache 0.00000119 0.00028351 - layer.3.k_cache 0.00136607 0.18994255 - layer.3.v_cache 0.00000217 0.00045241 - layer.4.k_cache 0.00338641 0.42049680 - layer.4.v_cache 0.00000303 0.00073137 - layer.4.output 0.00020569 0.03569448 - ------------------------------------------------------------------------------------- - TOTAL 0.00275548 0.38371587 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 742116 -BPFP 2.6891 bits/point -EBPFP 5.3783 equivalent bits/point -MSE 0.383716 ----------------------- -------------------------------------------------------- -Time: 3.279s Load: 0.009s, Pack+Encode: 1.891s, Decode+Unpack: 1.380s ----------------------- -------------------------------------------------------- -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 0.3837 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample101-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample101-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample102-layer4-item1.zst (5/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample102-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 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, 150, 128) -Output shape: (1, 150, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.output: torch.Size([1, 150, 4096]) -> torch.Size([1, 1, 150, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,332B, BPFP=1.1110 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 64,960B, BPFP=3.3833 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,540B, BPFP=2.1635 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 68,808B, BPFP=3.5838 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,504B, BPFP=2.6304 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 70,612B, BPFP=3.6777 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,216B, BPFP=2.6154 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,500B, BPFP=3.5677 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,292B, BPFP=2.2027 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,776B, BPFP=3.7383 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 200,664B, BPFP=2.6128 -⌛️ [2/4] FRONTEND: Frontend time: 1.868s (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, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.386s - -[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, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02593268 4.23715739 - layer.0.v_cache 0.00000026 0.00005952 - layer.1.k_cache 0.00305171 0.32987213 - layer.1.v_cache 0.00000092 0.00021761 - layer.2.k_cache 0.00119162 0.16194255 - layer.2.v_cache 0.00000123 0.00031155 - layer.3.k_cache 0.00128981 0.17507369 - layer.3.v_cache 0.00000247 0.00049698 - layer.4.k_cache 0.00345981 0.39989100 - layer.4.v_cache 0.00000333 0.00076049 - layer.4.output 0.00015266 0.03823782 - ------------------------------------------------------------------------------------- - TOTAL 0.00253889 0.38990959 - (elements=2,150,400) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2150400 -Total Bytes 751204 -BPFP 2.7947 bits/point -EBPFP 5.5893 equivalent bits/point -MSE 0.389910 ----------------------- -------------------------------------------------------- -Time: 3.263s Load: 0.010s, Pack+Encode: 1.868s, Decode+Unpack: 1.386s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3899 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample102-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample102-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample103-layer4-item1.zst (6/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample103-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 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, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,100B, BPFP=1.0216 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,248B, BPFP=3.1550 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,420B, BPFP=2.0997 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,208B, BPFP=3.3380 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,680B, BPFP=2.4815 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,848B, BPFP=3.3676 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,560B, BPFP=2.4760 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,444B, BPFP=3.3027 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,212B, BPFP=2.1363 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,720B, BPFP=3.4079 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 191,192B, BPFP=2.2096 -⌛️ [2/4] FRONTEND: Frontend time: 1.873s (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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.386s - -[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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02810763 4.68331494 - layer.0.v_cache 0.00000027 0.00006126 - layer.1.k_cache 0.00304808 0.36204795 - layer.1.v_cache 0.00000079 0.00019996 - layer.2.k_cache 0.00115532 0.16011069 - layer.2.v_cache 0.00000117 0.00028619 - layer.3.k_cache 0.00130126 0.17654320 - layer.3.v_cache 0.00000227 0.00045374 - layer.4.k_cache 0.00349612 0.43546909 - layer.4.v_cache 0.00000313 0.00075063 - layer.4.output 0.00017973 0.02897236 - ------------------------------------------------------------------------------------- - TOTAL 0.00270250 0.42393765 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 770632 -BPFP 2.5446 bits/point -EBPFP 5.0892 equivalent bits/point -MSE 0.423938 ----------------------- -------------------------------------------------------- -Time: 3.268s Load: 0.009s, Pack+Encode: 1.873s, Decode+Unpack: 1.386s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4239 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample103-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample103-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample105-layer4-item1.zst (7/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample105-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 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, 150, 128) -Output shape: (1, 150, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.output: torch.Size([1, 150, 4096]) -> torch.Size([1, 1, 150, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 20,916B, BPFP=1.0894 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 64,776B, BPFP=3.3737 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,140B, BPFP=2.1427 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 68,760B, BPFP=3.5812 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 49,328B, BPFP=2.5692 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 70,028B, BPFP=3.6473 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 49,216B, BPFP=2.5633 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 67,348B, BPFP=3.5077 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 41,196B, BPFP=2.1456 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 69,928B, BPFP=3.6421 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 182,896B, BPFP=2.3815 -⌛️ [2/4] FRONTEND: Frontend time: 1.879s (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, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.389s - -[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, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02945780 4.50196045 - layer.0.v_cache 0.00000027 0.00005955 - layer.1.k_cache 0.00309326 0.32338379 - layer.1.v_cache 0.00000086 0.00019771 - layer.2.k_cache 0.00115289 0.16346505 - layer.2.v_cache 0.00000113 0.00027487 - layer.3.k_cache 0.00132321 0.17818558 - layer.3.v_cache 0.00000214 0.00043841 - layer.4.k_cache 0.00338825 0.39726390 - layer.4.v_cache 0.00000324 0.00075191 - layer.4.output 0.00014331 0.02923849 - ------------------------------------------------------------------------------------- - TOTAL 0.00278545 0.40592394 - (elements=2,150,400) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2150400 -Total Bytes 725532 -BPFP 2.6992 bits/point -EBPFP 5.3983 equivalent bits/point -MSE 0.405924 ----------------------- -------------------------------------------------------- -Time: 3.278s Load: 0.010s, Pack+Encode: 1.879s, Decode+Unpack: 1.389s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4059 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample105-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample105-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample106-layer4-item1.zst (8/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample106-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: 0.009s - ------------------------------------------------------------- -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: 23,196B, BPFP=1.0181 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,400B, BPFP=2.9582 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,328B, BPFP=2.0334 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,344B, BPFP=3.2191 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,092B, BPFP=2.3741 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,804B, BPFP=3.1954 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,840B, BPFP=2.4070 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,344B, BPFP=3.1313 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,820B, BPFP=2.0550 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,220B, BPFP=3.2137 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 209,392B, BPFP=2.2976 -⌛️ [2/4] FRONTEND: Frontend time: 1.963s (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: 1.561s - -[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 0.02793351 4.39722983 - layer.0.v_cache 0.00000028 0.00006147 - layer.1.k_cache 0.00336537 0.40132244 - layer.1.v_cache 0.00000081 0.00018848 - layer.2.k_cache 0.00115814 0.16260715 - layer.2.v_cache 0.00000111 0.00026415 - layer.3.k_cache 0.00131051 0.17706388 - layer.3.v_cache 0.00000213 0.00042721 - layer.4.k_cache 0.00357284 0.45172573 - layer.4.v_cache 0.00000305 0.00068121 - layer.4.output 0.00019607 0.02867043 - ------------------------------------------------------------------------------------- - TOTAL 0.00272372 0.40758952 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 792780 -BPFP 2.4854 bits/point -EBPFP 4.9708 equivalent bits/point -MSE 0.407590 ----------------------- -------------------------------------------------------- -Time: 3.533s Load: 0.009s, Pack+Encode: 1.963s, Decode+Unpack: 1.561s ----------------------- -------------------------------------------------------- -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 0.4076 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample106-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample106-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample11-layer4-item1.zst (9/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample11-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 217, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 217, 128) -Output shape: (1, 217, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.0.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.1.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.1.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.2.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.2.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.3.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.3.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.4.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.4.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.4.output: torch.Size([1, 217, 4096]) -> torch.Size([1, 1, 217, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 29,464B, BPFP=1.0608 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 88,092B, BPFP=3.1715 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 56,876B, BPFP=2.0477 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 93,064B, BPFP=3.3505 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 67,640B, BPFP=2.4352 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 95,188B, BPFP=3.4270 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 68,648B, BPFP=2.4715 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 92,452B, BPFP=3.3285 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 57,444B, BPFP=2.0681 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 94,760B, BPFP=3.4116 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 290,072B, BPFP=2.6108 -⌛️ [2/4] FRONTEND: Frontend time: 2.566s (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, 217, 128]) - layer.0.v_cache: torch.Size([1, 8, 217, 128]) - layer.1.k_cache: torch.Size([1, 8, 217, 128]) - layer.1.v_cache: torch.Size([1, 8, 217, 128]) - layer.2.k_cache: torch.Size([1, 8, 217, 128]) - layer.2.v_cache: torch.Size([1, 8, 217, 128]) - layer.3.k_cache: torch.Size([1, 8, 217, 128]) - layer.3.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.k_cache: torch.Size([1, 8, 217, 128]) - layer.4.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.output: torch.Size([1, 217, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.916s - -[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, 217, 128]) - layer.0.v_cache: torch.Size([1, 8, 217, 128]) - layer.1.k_cache: torch.Size([1, 8, 217, 128]) - layer.1.v_cache: torch.Size([1, 8, 217, 128]) - layer.2.k_cache: torch.Size([1, 8, 217, 128]) - layer.2.v_cache: torch.Size([1, 8, 217, 128]) - layer.3.k_cache: torch.Size([1, 8, 217, 128]) - layer.3.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.k_cache: torch.Size([1, 8, 217, 128]) - layer.4.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.output: torch.Size([1, 217, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02570072 3.91265532 - layer.0.v_cache 0.00000028 0.00005582 - layer.1.k_cache 0.00315590 0.34952514 - layer.1.v_cache 0.00000089 0.00019544 - layer.2.k_cache 0.00124266 0.16541487 - layer.2.v_cache 0.00000119 0.00027144 - layer.3.k_cache 0.00129581 0.17318777 - layer.3.v_cache 0.00000363 0.00049655 - layer.4.k_cache 0.00355111 0.41578491 - layer.4.v_cache 0.00000379 0.00073507 - layer.4.output 0.00017838 0.03267589 - ------------------------------------------------------------------------------------- - TOTAL 0.00254782 0.36778756 - (elements=3,110,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3110912 -Total Bytes 1033700 -BPFP 2.6583 bits/point -EBPFP 5.3165 equivalent bits/point -MSE 0.367788 ----------------------- -------------------------------------------------------- -Time: 4.496s Load: 0.013s, Pack+Encode: 2.566s, Decode+Unpack: 1.916s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 217, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3678 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample11-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample11-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample110-layer4-item1.zst (10/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample110-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: 23,864B, BPFP=1.0777 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,108B, BPFP=3.0757 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,036B, BPFP=2.0789 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,076B, BPFP=3.3000 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,448B, BPFP=2.4588 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 77,648B, BPFP=3.5065 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,632B, BPFP=2.4671 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,788B, BPFP=3.2419 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,992B, BPFP=2.1221 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,568B, BPFP=3.3223 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 223,872B, BPFP=2.5275 -⌛️ [2/4] FRONTEND: Frontend time: 1.939s (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: 1.476s - -[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 0.02750326 4.59508678 - layer.0.v_cache 0.00000027 0.00006299 - layer.1.k_cache 0.00310050 0.39798348 - layer.1.v_cache 0.00000106 0.00020604 - layer.2.k_cache 0.00119927 0.16897994 - layer.2.v_cache 0.00000121 0.00029556 - layer.3.k_cache 0.00131341 0.17670507 - layer.3.v_cache 0.00000232 0.00046130 - layer.4.k_cache 0.00354461 0.45955407 - layer.4.v_cache 0.00000318 0.00072739 - layer.4.output 0.00022996 0.03691980 - ------------------------------------------------------------------------------------- - TOTAL 0.00268492 0.42483870 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 814032 -BPFP 2.6258 bits/point -EBPFP 5.2515 equivalent bits/point -MSE 0.424839 ----------------------- -------------------------------------------------------- -Time: 3.425s Load: 0.011s, Pack+Encode: 1.939s, Decode+Unpack: 1.476s ----------------------- -------------------------------------------------------- -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 0.4248 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample110-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample110-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample111-layer4-item1.zst (11/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample111-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: 23,816B, BPFP=1.0453 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,884B, BPFP=2.9795 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,668B, BPFP=2.0483 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 75,372B, BPFP=3.3081 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,060B, BPFP=2.3727 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,360B, BPFP=3.2637 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,836B, BPFP=2.4068 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,932B, BPFP=3.2010 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,960B, BPFP=2.0611 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,632B, BPFP=3.2756 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 234,304B, BPFP=2.5709 -⌛️ [2/4] FRONTEND: Frontend time: 1.868s (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: 1.388s - -[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, 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 0.02635046 4.44456225 - layer.0.v_cache 0.00000029 0.00006485 - layer.1.k_cache 0.00306277 0.41434925 - layer.1.v_cache 0.00000083 0.00020867 - layer.2.k_cache 0.00121657 0.16483877 - layer.2.v_cache 0.00000121 0.00028215 - layer.3.k_cache 0.00128658 0.17298636 - layer.3.v_cache 0.00000219 0.00045112 - layer.4.k_cache 0.00351850 0.45197262 - layer.4.v_cache 0.00000322 0.00075458 - layer.4.output 0.00018894 0.03624574 - ------------------------------------------------------------------------------------- - TOTAL 0.00258560 0.41396097 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 825824 -BPFP 2.5890 bits/point -EBPFP 5.1780 equivalent bits/point -MSE 0.413961 ----------------------- -------------------------------------------------------- -Time: 3.267s Load: 0.010s, Pack+Encode: 1.868s, Decode+Unpack: 1.388s ----------------------- -------------------------------------------------------- -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 0.4140 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample111-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample111-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample112-layer4-item1.zst (12/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample112-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 164, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 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, 164, 128) -Output shape: (1, 164, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.0.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.1.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.1.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.2.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.2.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.3.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.3.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.4.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.4.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.4.output: torch.Size([1, 164, 4096]) -> torch.Size([1, 1, 164, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,116B, BPFP=1.0535 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,176B, BPFP=3.2001 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,780B, BPFP=2.1332 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,228B, BPFP=3.3931 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 52,716B, BPFP=2.5112 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,948B, BPFP=3.4274 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 52,924B, BPFP=2.5212 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,104B, BPFP=3.3872 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,752B, BPFP=2.1795 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,904B, BPFP=3.6159 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 188,068B, BPFP=2.2398 -⌛️ [2/4] FRONTEND: Frontend time: 1.861s (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, 164, 128]) - layer.0.v_cache: torch.Size([1, 8, 164, 128]) - layer.1.k_cache: torch.Size([1, 8, 164, 128]) - layer.1.v_cache: torch.Size([1, 8, 164, 128]) - layer.2.k_cache: torch.Size([1, 8, 164, 128]) - layer.2.v_cache: torch.Size([1, 8, 164, 128]) - layer.3.k_cache: torch.Size([1, 8, 164, 128]) - layer.3.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.k_cache: torch.Size([1, 8, 164, 128]) - layer.4.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.output: torch.Size([1, 164, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.390s - -[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, 164, 128]) - layer.0.v_cache: torch.Size([1, 8, 164, 128]) - layer.1.k_cache: torch.Size([1, 8, 164, 128]) - layer.1.v_cache: torch.Size([1, 8, 164, 128]) - layer.2.k_cache: torch.Size([1, 8, 164, 128]) - layer.2.v_cache: torch.Size([1, 8, 164, 128]) - layer.3.k_cache: torch.Size([1, 8, 164, 128]) - layer.3.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.k_cache: torch.Size([1, 8, 164, 128]) - layer.4.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.output: torch.Size([1, 164, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02515400 4.28066942 - layer.0.v_cache 0.00000027 0.00005829 - layer.1.k_cache 0.00312787 0.35866726 - layer.1.v_cache 0.00000077 0.00018756 - layer.2.k_cache 0.00115327 0.16244802 - layer.2.v_cache 0.00000115 0.00027044 - layer.3.k_cache 0.00131309 0.17813113 - layer.3.v_cache 0.00000211 0.00041750 - layer.4.k_cache 0.00347769 0.42693864 - layer.4.v_cache 0.00000342 0.00076073 - layer.4.output 0.00018780 0.03407852 - ------------------------------------------------------------------------------------- - TOTAL 0.00249892 0.39606165 - (elements=2,351,104) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2351104 -Total Bytes 763716 -BPFP 2.5987 bits/point -EBPFP 5.1973 equivalent bits/point -MSE 0.396062 ----------------------- -------------------------------------------------------- -Time: 3.259s Load: 0.009s, Pack+Encode: 1.861s, Decode+Unpack: 1.390s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 164, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3961 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample112-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample112-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample114-layer4-item1.zst (13/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample114-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 144, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 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, 144, 128) -Output shape: (1, 144, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.0.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.1.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.1.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.2.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.2.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.3.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.3.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.4.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.4.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.4.output: torch.Size([1, 144, 4096]) -> torch.Size([1, 1, 144, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 20,140B, BPFP=1.0927 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 64,364B, BPFP=3.4920 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 40,884B, BPFP=2.2181 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 68,512B, BPFP=3.7170 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 48,740B, BPFP=2.6443 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 69,068B, BPFP=3.7472 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 48,972B, BPFP=2.6569 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 67,172B, BPFP=3.6443 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 40,576B, BPFP=2.2014 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 69,760B, BPFP=3.7847 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 182,764B, BPFP=2.4789 -⌛️ [2/4] FRONTEND: Frontend time: 1.864s (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, 144, 128]) - layer.0.v_cache: torch.Size([1, 8, 144, 128]) - layer.1.k_cache: torch.Size([1, 8, 144, 128]) - layer.1.v_cache: torch.Size([1, 8, 144, 128]) - layer.2.k_cache: torch.Size([1, 8, 144, 128]) - layer.2.v_cache: torch.Size([1, 8, 144, 128]) - layer.3.k_cache: torch.Size([1, 8, 144, 128]) - layer.3.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.k_cache: torch.Size([1, 8, 144, 128]) - layer.4.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.output: torch.Size([1, 144, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.378s - -[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, 144, 128]) - layer.0.v_cache: torch.Size([1, 8, 144, 128]) - layer.1.k_cache: torch.Size([1, 8, 144, 128]) - layer.1.v_cache: torch.Size([1, 8, 144, 128]) - layer.2.k_cache: torch.Size([1, 8, 144, 128]) - layer.2.v_cache: torch.Size([1, 8, 144, 128]) - layer.3.k_cache: torch.Size([1, 8, 144, 128]) - layer.3.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.k_cache: torch.Size([1, 8, 144, 128]) - layer.4.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.output: torch.Size([1, 144, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02690511 4.06050025 - layer.0.v_cache 0.00000027 0.00005958 - layer.1.k_cache 0.00313749 0.31513879 - layer.1.v_cache 0.00000108 0.00021020 - layer.2.k_cache 0.00118318 0.15751741 - layer.2.v_cache 0.00000114 0.00027599 - layer.3.k_cache 0.00131682 0.17656350 - layer.3.v_cache 0.00000220 0.00046234 - layer.4.k_cache 0.00346269 0.36676685 - layer.4.v_cache 0.00000333 0.00075412 - layer.4.output 0.00016306 0.03565175 - ------------------------------------------------------------------------------------- - TOTAL 0.00261897 0.37291829 - (elements=2,064,384) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2064384 -Total Bytes 720952 -BPFP 2.7939 bits/point -EBPFP 5.5877 equivalent bits/point -MSE 0.372918 ----------------------- -------------------------------------------------------- -Time: 3.251s Load: 0.009s, Pack+Encode: 1.864s, Decode+Unpack: 1.378s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 144, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3729 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample114-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample114-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample116-layer4-item1.zst (14/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample116-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 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, 171, 128) -Output shape: (1, 171, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.output: torch.Size([1, 171, 4096]) -> torch.Size([1, 1, 171, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,544B, BPFP=1.0300 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,468B, BPFP=3.0824 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,672B, BPFP=2.1323 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,044B, BPFP=3.3829 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,768B, BPFP=2.4565 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,528B, BPFP=3.3136 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,228B, BPFP=2.4775 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,672B, BPFP=3.2745 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,692B, BPFP=2.1332 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,176B, BPFP=3.4346 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 195,420B, BPFP=2.2320 -⌛️ [2/4] FRONTEND: Frontend time: 1.872s (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, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.393s - -[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, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02584258 4.49618155 - layer.0.v_cache 0.00000026 0.00005885 - layer.1.k_cache 0.00320420 0.39934937 - layer.1.v_cache 0.00000080 0.00020137 - layer.2.k_cache 0.00119855 0.16198695 - layer.2.v_cache 0.00000108 0.00026018 - layer.3.k_cache 0.00129062 0.17273151 - layer.3.v_cache 0.00000212 0.00045377 - layer.4.k_cache 0.00351305 0.42009106 - layer.4.v_cache 0.00000315 0.00074326 - layer.4.output 0.00016406 0.02971871 - ------------------------------------------------------------------------------------- - TOTAL 0.00255090 0.41220948 - (elements=2,451,456) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2451456 -Total Bytes 780212 -BPFP 2.5461 bits/point -EBPFP 5.0922 equivalent bits/point -MSE 0.412209 ----------------------- -------------------------------------------------------- -Time: 3.274s Load: 0.009s, Pack+Encode: 1.872s, Decode+Unpack: 1.393s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4122 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample116-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample116-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample117-layer4-item1.zst (15/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample117-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 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, 175, 128) -Output shape: (1, 175, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.output: torch.Size([1, 175, 4096]) -> torch.Size([1, 1, 175, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,180B, BPFP=1.0348 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,732B, BPFP=3.0238 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,708B, BPFP=2.0852 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 78,412B, BPFP=3.5005 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,636B, BPFP=2.4391 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 78,776B, BPFP=3.5168 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,908B, BPFP=2.4512 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,948B, BPFP=3.2120 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,284B, BPFP=2.1109 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 76,064B, BPFP=3.3957 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 201,840B, BPFP=2.2527 -⌛️ [2/4] FRONTEND: Frontend time: 1.868s (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, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.396s - -[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, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02705980 4.48135045 - layer.0.v_cache 0.00000027 0.00006506 - layer.1.k_cache 0.00311821 0.41676801 - layer.1.v_cache 0.00000085 0.00019509 - layer.2.k_cache 0.00114782 0.16142798 - layer.2.v_cache 0.00000111 0.00027994 - layer.3.k_cache 0.00131325 0.18285466 - layer.3.v_cache 0.00000221 0.00045950 - layer.4.k_cache 0.00342930 0.44862544 - layer.4.v_cache 0.00000305 0.00075115 - layer.4.output 0.00022257 0.03387047 - ------------------------------------------------------------------------------------- - TOTAL 0.00264044 0.41630423 - (elements=2,508,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2508800 -Total Bytes 801488 -BPFP 2.5558 bits/point -EBPFP 5.1115 equivalent bits/point -MSE 0.416304 ----------------------- -------------------------------------------------------- -Time: 3.273s Load: 0.009s, Pack+Encode: 1.868s, Decode+Unpack: 1.396s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4163 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample117-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample117-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample12-layer4-item1.zst (16/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample12-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 218, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 218, 128) -Output shape: (1, 218, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.0.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.1.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.1.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.2.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.2.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.3.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.3.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.4.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.4.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.4.output: torch.Size([1, 218, 4096]) -> torch.Size([1, 1, 218, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 28,552B, BPFP=1.0232 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 88,908B, BPFP=3.1862 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 57,528B, BPFP=2.0616 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 93,988B, BPFP=3.3683 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 69,336B, BPFP=2.4848 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 94,604B, BPFP=3.3903 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 69,224B, BPFP=2.4808 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 92,748B, BPFP=3.3238 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 58,072B, BPFP=2.0811 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 96,044B, BPFP=3.4419 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 240,788B, BPFP=2.1573 -⌛️ [2/4] FRONTEND: Frontend time: 2.224s (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, 218, 128]) - layer.0.v_cache: torch.Size([1, 8, 218, 128]) - layer.1.k_cache: torch.Size([1, 8, 218, 128]) - layer.1.v_cache: torch.Size([1, 8, 218, 128]) - layer.2.k_cache: torch.Size([1, 8, 218, 128]) - layer.2.v_cache: torch.Size([1, 8, 218, 128]) - layer.3.k_cache: torch.Size([1, 8, 218, 128]) - layer.3.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.k_cache: torch.Size([1, 8, 218, 128]) - layer.4.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.output: torch.Size([1, 218, 4096]) -⌛️ [3/4] BACKEND: Backend time: 31.040s - -[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, 218, 128]) - layer.0.v_cache: torch.Size([1, 8, 218, 128]) - layer.1.k_cache: torch.Size([1, 8, 218, 128]) - layer.1.v_cache: torch.Size([1, 8, 218, 128]) - layer.2.k_cache: torch.Size([1, 8, 218, 128]) - layer.2.v_cache: torch.Size([1, 8, 218, 128]) - layer.3.k_cache: torch.Size([1, 8, 218, 128]) - layer.3.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.k_cache: torch.Size([1, 8, 218, 128]) - layer.4.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.output: torch.Size([1, 218, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02786819 3.90474820 - layer.0.v_cache 0.00000028 0.00005933 - layer.1.k_cache 0.00304396 0.34328195 - layer.1.v_cache 0.00000081 0.00020497 - layer.2.k_cache 0.00116087 0.16176399 - layer.2.v_cache 0.00000117 0.00027319 - layer.3.k_cache 0.00132673 0.18415427 - layer.3.v_cache 0.00000218 0.00045775 - layer.4.k_cache 0.00351200 0.41743672 - layer.4.v_cache 0.00000322 0.00072083 - layer.4.output 0.00018944 0.02976046 - ------------------------------------------------------------------------------------- - TOTAL 0.00269123 0.36658164 - (elements=3,125,248) ----------------------- --------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- --------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- --------------------------------------------------------- -Total Elements 3125248 -Total Bytes 989792 -BPFP 2.5337 bits/point -EBPFP 5.0673 equivalent bits/point -MSE 0.366582 ----------------------- --------------------------------------------------------- -Time: 33.276s Load: 0.012s, Pack+Encode: 2.224s, Decode+Unpack: 31.040s ----------------------- --------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 218, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3666 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample12-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample12-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample120-layer4-item1.zst (17/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample120-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 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, 171, 128) -Output shape: (1, 171, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.output: torch.Size([1, 171, 4096]) -> torch.Size([1, 1, 171, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,040B, BPFP=1.0526 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,064B, BPFP=3.1096 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,076B, BPFP=2.1051 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,444B, BPFP=3.3554 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,812B, BPFP=2.4585 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,412B, BPFP=3.3540 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,080B, BPFP=2.4708 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,844B, BPFP=3.2367 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,152B, BPFP=2.1086 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,820B, BPFP=3.3269 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 192,432B, BPFP=2.1979 -⌛️ [2/4] FRONTEND: Frontend time: 2.000s (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, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.589s - -[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, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02655128 4.19472035 - layer.0.v_cache 0.00000026 0.00005696 - layer.1.k_cache 0.00304961 0.39006150 - layer.1.v_cache 0.00000080 0.00019003 - layer.2.k_cache 0.00119340 0.15407672 - layer.2.v_cache 0.00000107 0.00025712 - layer.3.k_cache 0.00127930 0.17385115 - layer.3.v_cache 0.00000211 0.00041211 - layer.4.k_cache 0.00352197 0.43315553 - layer.4.v_cache 0.00000299 0.00068931 - layer.4.output 0.00016153 0.03124869 - ------------------------------------------------------------------------------------- - TOTAL 0.00258921 0.39089040 - (elements=2,451,456) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2451456 -Total Bytes 774176 -BPFP 2.5264 bits/point -EBPFP 5.0528 equivalent bits/point -MSE 0.390890 ----------------------- -------------------------------------------------------- -Time: 3.599s Load: 0.010s, Pack+Encode: 2.000s, Decode+Unpack: 1.589s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3909 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample120-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample120-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample122-layer4-item1.zst (18/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample122-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 190, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 190, 128) -Output shape: (1, 190, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.0.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.1.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.1.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.2.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.2.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.3.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.3.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.4.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.4.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.4.output: torch.Size([1, 190, 4096]) -> torch.Size([1, 1, 190, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 26,516B, BPFP=1.0903 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,728B, BPFP=2.7437 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,300B, BPFP=1.8627 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,712B, BPFP=2.9898 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,164B, BPFP=2.1860 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,688B, BPFP=3.0299 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,448B, BPFP=2.2388 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,416B, BPFP=2.9365 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,272B, BPFP=1.9026 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,620B, BPFP=3.0683 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 212,764B, BPFP=2.1871 -⌛️ [2/4] FRONTEND: Frontend time: 1.945s (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, 190, 128]) - layer.0.v_cache: torch.Size([1, 8, 190, 128]) - layer.1.k_cache: torch.Size([1, 8, 190, 128]) - layer.1.v_cache: torch.Size([1, 8, 190, 128]) - layer.2.k_cache: torch.Size([1, 8, 190, 128]) - layer.2.v_cache: torch.Size([1, 8, 190, 128]) - layer.3.k_cache: torch.Size([1, 8, 190, 128]) - layer.3.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.k_cache: torch.Size([1, 8, 190, 128]) - layer.4.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.output: torch.Size([1, 190, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.555s - -[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, 190, 128]) - layer.0.v_cache: torch.Size([1, 8, 190, 128]) - layer.1.k_cache: torch.Size([1, 8, 190, 128]) - layer.1.v_cache: torch.Size([1, 8, 190, 128]) - layer.2.k_cache: torch.Size([1, 8, 190, 128]) - layer.2.v_cache: torch.Size([1, 8, 190, 128]) - layer.3.k_cache: torch.Size([1, 8, 190, 128]) - layer.3.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.k_cache: torch.Size([1, 8, 190, 128]) - layer.4.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.output: torch.Size([1, 190, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02682430 4.77272403 - layer.0.v_cache 0.00000028 0.00005931 - layer.1.k_cache 0.00318839 0.37421000 - layer.1.v_cache 0.00000076 0.00016799 - layer.2.k_cache 0.00117212 0.16995219 - layer.2.v_cache 0.00000110 0.00023325 - layer.3.k_cache 0.00129629 0.17404006 - layer.3.v_cache 0.00000212 0.00039449 - layer.4.k_cache 0.00366745 0.48572753 - layer.4.v_cache 0.00000296 0.00060340 - layer.4.output 0.00018717 0.03099911 - ------------------------------------------------------------------------------------- - TOTAL 0.00263603 0.43586491 - (elements=2,723,840) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2723840 -Total Bytes 797628 -BPFP 2.3427 bits/point -EBPFP 4.6853 equivalent bits/point -MSE 0.435865 ----------------------- -------------------------------------------------------- -Time: 3.514s Load: 0.013s, Pack+Encode: 1.945s, Decode+Unpack: 1.555s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 190, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4359 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample122-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample122-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample125-layer4-item1.zst (19/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample125-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 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, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,468B, BPFP=1.0205 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,352B, BPFP=3.0592 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,716B, BPFP=2.0765 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,672B, BPFP=3.3009 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,304B, BPFP=2.4211 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,528B, BPFP=3.2943 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,104B, BPFP=2.4575 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,720B, BPFP=3.2122 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,160B, BPFP=2.0967 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,540B, BPFP=3.3403 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 201,464B, BPFP=2.2877 -⌛️ [2/4] FRONTEND: Frontend time: 1.980s (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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.522s - -[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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02742122 4.32678010 - layer.0.v_cache 0.00000027 0.00006076 - layer.1.k_cache 0.00311153 0.39909824 - layer.1.v_cache 0.00000078 0.00018777 - layer.2.k_cache 0.00118723 0.16392971 - layer.2.v_cache 0.00000107 0.00025556 - layer.3.k_cache 0.00131758 0.17533262 - layer.3.v_cache 0.00000212 0.00042026 - layer.4.k_cache 0.00355508 0.44118340 - layer.4.v_cache 0.00000305 0.00068752 - layer.4.output 0.00016643 0.03099265 - ------------------------------------------------------------------------------------- - TOTAL 0.00266183 0.40227904 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 780028 -BPFP 2.5307 bits/point -EBPFP 5.0614 equivalent bits/point -MSE 0.402279 ----------------------- -------------------------------------------------------- -Time: 3.511s Load: 0.010s, Pack+Encode: 1.980s, Decode+Unpack: 1.522s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4023 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample125-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample125-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample126-layer4-item1.zst (20/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample126-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 165, 128) -Output shape: (1, 165, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.output: torch.Size([1, 165, 4096]) -> torch.Size([1, 1, 165, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,460B, BPFP=1.0634 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 69,100B, BPFP=3.2718 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,640B, BPFP=2.1136 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,216B, BPFP=3.4193 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,312B, BPFP=2.5242 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,680B, BPFP=3.4413 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,448B, BPFP=2.5307 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,708B, BPFP=3.3953 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,744B, BPFP=2.1659 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,784B, BPFP=3.4462 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 220,336B, BPFP=2.6081 -⌛️ [2/4] FRONTEND: Frontend time: 1.887s (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, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.391s - -[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, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02563144 4.62748284 - layer.0.v_cache 0.00000028 0.00006613 - layer.1.k_cache 0.00322103 0.34400392 - layer.1.v_cache 0.00000086 0.00020605 - layer.2.k_cache 0.00118144 0.16566950 - layer.2.v_cache 0.00000111 0.00027537 - layer.3.k_cache 0.00130833 0.17757828 - layer.3.v_cache 0.00000230 0.00047733 - layer.4.k_cache 0.00357125 0.47757837 - layer.4.v_cache 0.00000311 0.00072274 - layer.4.output 0.00020081 0.03552419 - ------------------------------------------------------------------------------------- - TOTAL 0.00255174 0.42401123 - (elements=2,365,440) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2365440 -Total Bytes 798428 -BPFP 2.7003 bits/point -EBPFP 5.4006 equivalent bits/point -MSE 0.424011 ----------------------- -------------------------------------------------------- -Time: 3.290s Load: 0.012s, Pack+Encode: 1.887s, Decode+Unpack: 1.391s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4240 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample126-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample126-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample129-layer4-item1.zst (21/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample129-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 174, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 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, 174, 128) -Output shape: (1, 174, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.0.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.1.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.1.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.2.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.2.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.3.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.3.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.4.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.4.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.4.output: torch.Size([1, 174, 4096]) -> torch.Size([1, 1, 174, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,036B, BPFP=1.0343 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,024B, BPFP=3.0093 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,964B, BPFP=2.0638 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,500B, BPFP=3.2552 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,260B, BPFP=2.3913 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,908B, BPFP=3.2286 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,900B, BPFP=2.4201 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,792B, BPFP=3.1785 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,476B, BPFP=2.0867 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,724B, BPFP=3.3102 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 197,980B, BPFP=2.2223 -⌛️ [2/4] FRONTEND: Frontend time: 1.855s (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, 174, 128]) - layer.0.v_cache: torch.Size([1, 8, 174, 128]) - layer.1.k_cache: torch.Size([1, 8, 174, 128]) - layer.1.v_cache: torch.Size([1, 8, 174, 128]) - layer.2.k_cache: torch.Size([1, 8, 174, 128]) - layer.2.v_cache: torch.Size([1, 8, 174, 128]) - layer.3.k_cache: torch.Size([1, 8, 174, 128]) - layer.3.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.k_cache: torch.Size([1, 8, 174, 128]) - layer.4.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.output: torch.Size([1, 174, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.384s - -[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, 174, 128]) - layer.0.v_cache: torch.Size([1, 8, 174, 128]) - layer.1.k_cache: torch.Size([1, 8, 174, 128]) - layer.1.v_cache: torch.Size([1, 8, 174, 128]) - layer.2.k_cache: torch.Size([1, 8, 174, 128]) - layer.2.v_cache: torch.Size([1, 8, 174, 128]) - layer.3.k_cache: torch.Size([1, 8, 174, 128]) - layer.3.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.k_cache: torch.Size([1, 8, 174, 128]) - layer.4.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.output: torch.Size([1, 174, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02884222 4.69317206 - layer.0.v_cache 0.00000026 0.00006044 - layer.1.k_cache 0.00305335 0.41214932 - layer.1.v_cache 0.00000075 0.00018480 - layer.2.k_cache 0.00114914 0.16592335 - layer.2.v_cache 0.00000112 0.00026307 - layer.3.k_cache 0.00130410 0.17907532 - layer.3.v_cache 0.00000204 0.00042726 - layer.4.k_cache 0.00362435 0.44587580 - layer.4.v_cache 0.00000301 0.00068368 - layer.4.output 0.00017507 0.03450521 - ------------------------------------------------------------------------------------- - TOTAL 0.00276290 0.43113114 - (elements=2,494,464) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2494464 -Total Bytes 776564 -BPFP 2.4905 bits/point -EBPFP 4.9810 equivalent bits/point -MSE 0.431131 ----------------------- -------------------------------------------------------- -Time: 3.249s Load: 0.009s, Pack+Encode: 1.855s, Decode+Unpack: 1.384s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 174, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4311 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample129-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample129-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample130-layer4-item1.zst (22/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample130-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 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, 152, 128) -Output shape: (1, 152, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.output: torch.Size([1, 152, 4096]) -> torch.Size([1, 1, 152, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,312B, BPFP=1.0954 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 65,720B, BPFP=3.3779 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,620B, BPFP=2.1392 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 69,652B, BPFP=3.5800 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,604B, BPFP=2.6009 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 70,768B, BPFP=3.6373 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,388B, BPFP=2.5898 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,504B, BPFP=3.5210 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 41,912B, BPFP=2.1542 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,540B, BPFP=3.6770 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 191,040B, BPFP=2.4548 -⌛️ [2/4] FRONTEND: Frontend time: 1.867s (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, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.445s - -[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, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02732493 3.86800625 - layer.0.v_cache 0.00000027 0.00006008 - layer.1.k_cache 0.00311055 0.32227082 - layer.1.v_cache 0.00000090 0.00020234 - layer.2.k_cache 0.00122034 0.16098224 - layer.2.v_cache 0.00000121 0.00028434 - layer.3.k_cache 0.00130561 0.17791653 - layer.3.v_cache 0.00000231 0.00046288 - layer.4.k_cache 0.00351005 0.40600892 - layer.4.v_cache 0.00000314 0.00071098 - layer.4.output 0.00016281 0.03495307 - ------------------------------------------------------------------------------------- - TOTAL 0.00265218 0.36262269 - (elements=2,179,072) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2179072 -Total Bytes 743060 -BPFP 2.7280 bits/point -EBPFP 5.4560 equivalent bits/point -MSE 0.362623 ----------------------- -------------------------------------------------------- -Time: 3.321s Load: 0.010s, Pack+Encode: 1.867s, Decode+Unpack: 1.445s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3626 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample130-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample130-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample144-layer4-item1.zst (23/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample144-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: 20,732B, BPFP=1.0517 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,680B, BPFP=3.3827 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,228B, BPFP=2.0915 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,020B, BPFP=3.5522 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 49,328B, BPFP=2.5024 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,224B, BPFP=3.6640 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,168B, BPFP=2.5450 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,416B, BPFP=3.4708 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 41,804B, BPFP=2.1207 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,112B, BPFP=3.6075 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 187,052B, BPFP=2.3723 -⌛️ [2/4] FRONTEND: Frontend time: 1.968s (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: 1.451s - -[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 0.02760743 4.21933261 - layer.0.v_cache 0.00000026 0.00006026 - layer.1.k_cache 0.00317114 0.33776905 - layer.1.v_cache 0.00000086 0.00019431 - layer.2.k_cache 0.00117006 0.16455151 - layer.2.v_cache 0.00000109 0.00026520 - layer.3.k_cache 0.00134568 0.18150554 - layer.3.v_cache 0.00000205 0.00044172 - layer.4.k_cache 0.00345514 0.43527687 - layer.4.v_cache 0.00000304 0.00073385 - layer.4.output 0.00019092 0.03888336 - ------------------------------------------------------------------------------------- - TOTAL 0.00268003 0.39254745 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 738764 -BPFP 2.6770 bits/point -EBPFP 5.3540 equivalent bits/point -MSE 0.392547 ----------------------- -------------------------------------------------------- -Time: 3.428s Load: 0.008s, Pack+Encode: 1.968s, Decode+Unpack: 1.451s ----------------------- -------------------------------------------------------- -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 0.3925 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample144-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample144-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample147-layer4-item1.zst (24/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample147-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 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, 165, 128) -Output shape: (1, 165, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.output: torch.Size([1, 165, 4096]) -> torch.Size([1, 1, 165, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,876B, BPFP=1.0358 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,596B, BPFP=3.2006 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,356B, BPFP=2.1002 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,584B, BPFP=3.3894 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 52,400B, BPFP=2.4811 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,040B, BPFP=3.4110 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,568B, BPFP=2.5364 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,400B, BPFP=3.3333 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,728B, BPFP=2.1652 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,624B, BPFP=3.4386 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 192,724B, BPFP=2.2813 -⌛️ [2/4] FRONTEND: Frontend time: 1.874s (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, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.383s - -[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, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02663295 4.25289344 - layer.0.v_cache 0.00000027 0.00006156 - layer.1.k_cache 0.00314363 0.36252860 - layer.1.v_cache 0.00000082 0.00020407 - layer.2.k_cache 0.00118021 0.16515461 - layer.2.v_cache 0.00000116 0.00028049 - layer.3.k_cache 0.00128344 0.17672726 - layer.3.v_cache 0.00000214 0.00044097 - layer.4.k_cache 0.00361995 0.46983250 - layer.4.v_cache 0.00000298 0.00071824 - layer.4.output 0.00016477 0.03497472 - ------------------------------------------------------------------------------------- - TOTAL 0.00260904 0.39776719 - (elements=2,365,440) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2365440 -Total Bytes 764896 -BPFP 2.5869 bits/point -EBPFP 5.1738 equivalent bits/point -MSE 0.397767 ----------------------- -------------------------------------------------------- -Time: 3.267s Load: 0.010s, Pack+Encode: 1.874s, Decode+Unpack: 1.383s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3978 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample147-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample147-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample15-layer4-item1.zst (25/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample15-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 209, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 209, 128) -Output shape: (1, 209, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.0.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.1.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.1.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.2.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.2.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.3.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.3.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.4.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.4.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.4.output: torch.Size([1, 209, 4096]) -> torch.Size([1, 1, 209, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 26,908B, BPFP=1.0058 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 85,660B, BPFP=3.2020 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 55,308B, BPFP=2.0674 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 90,844B, BPFP=3.3958 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 66,256B, BPFP=2.4767 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 91,084B, BPFP=3.4048 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 66,360B, BPFP=2.4806 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 89,764B, BPFP=3.3554 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 55,080B, BPFP=2.0589 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 92,972B, BPFP=3.4753 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 233,576B, BPFP=2.1828 -⌛️ [2/4] FRONTEND: Frontend time: 16.194s (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, 209, 128]) - layer.0.v_cache: torch.Size([1, 8, 209, 128]) - layer.1.k_cache: torch.Size([1, 8, 209, 128]) - layer.1.v_cache: torch.Size([1, 8, 209, 128]) - layer.2.k_cache: torch.Size([1, 8, 209, 128]) - layer.2.v_cache: torch.Size([1, 8, 209, 128]) - layer.3.k_cache: torch.Size([1, 8, 209, 128]) - layer.3.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.k_cache: torch.Size([1, 8, 209, 128]) - layer.4.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.output: torch.Size([1, 209, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.969s - -[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, 209, 128]) - layer.0.v_cache: torch.Size([1, 8, 209, 128]) - layer.1.k_cache: torch.Size([1, 8, 209, 128]) - layer.1.v_cache: torch.Size([1, 8, 209, 128]) - layer.2.k_cache: torch.Size([1, 8, 209, 128]) - layer.2.v_cache: torch.Size([1, 8, 209, 128]) - layer.3.k_cache: torch.Size([1, 8, 209, 128]) - layer.3.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.k_cache: torch.Size([1, 8, 209, 128]) - layer.4.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.output: torch.Size([1, 209, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02572924 4.37615091 - layer.0.v_cache 0.00000027 0.00005862 - layer.1.k_cache 0.00299828 0.33944465 - layer.1.v_cache 0.00000076 0.00017961 - layer.2.k_cache 0.00126581 0.16171913 - layer.2.v_cache 0.00000114 0.00026182 - layer.3.k_cache 0.00131754 0.17583834 - layer.3.v_cache 0.00000208 0.00042776 - layer.4.k_cache 0.00354781 0.41432500 - layer.4.v_cache 0.00000302 0.00066753 - layer.4.output 0.00016605 0.02807737 - ------------------------------------------------------------------------------------- - TOTAL 0.00253787 0.39867020 - (elements=2,996,224) ----------------------- --------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- --------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- --------------------------------------------------------- -Total Elements 2996224 -Total Bytes 953812 -BPFP 2.5467 bits/point -EBPFP 5.0934 equivalent bits/point -MSE 0.398670 ----------------------- --------------------------------------------------------- -Time: 18.175s Load: 0.012s, Pack+Encode: 16.194s, Decode+Unpack: 1.969s ----------------------- --------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 209, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3987 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample15-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample15-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample152-layer4-item1.zst (26/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample152-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 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, 171, 128) -Output shape: (1, 171, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.output: torch.Size([1, 171, 4096]) -> torch.Size([1, 1, 171, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,764B, BPFP=1.1314 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,128B, BPFP=3.1126 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,336B, BPFP=2.1170 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,932B, BPFP=3.3321 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,776B, BPFP=2.4569 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,248B, BPFP=3.3465 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,388B, BPFP=2.4848 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,292B, BPFP=3.2571 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,268B, BPFP=2.1139 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,276B, BPFP=3.3935 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 216,708B, BPFP=2.4752 -⌛️ [2/4] FRONTEND: Frontend time: 1.987s (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, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.561s - -[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, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02710774 5.15803787 - layer.0.v_cache 0.00000028 0.00006555 - layer.1.k_cache 0.00313176 0.41794653 - layer.1.v_cache 0.00000087 0.00021155 - layer.2.k_cache 0.00119500 0.16344944 - layer.2.v_cache 0.00000129 0.00028519 - layer.3.k_cache 0.00128502 0.17446389 - layer.3.v_cache 0.00000223 0.00045550 - layer.4.k_cache 0.00358421 0.43693186 - layer.4.v_cache 0.00000308 0.00072484 - layer.4.output 0.00020665 0.03300124 - ------------------------------------------------------------------------------------- - TOTAL 0.00265272 0.46318408 - (elements=2,451,456) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2451456 -Total Bytes 802116 -BPFP 2.6176 bits/point -EBPFP 5.2352 equivalent bits/point -MSE 0.463184 ----------------------- -------------------------------------------------------- -Time: 3.556s Load: 0.009s, Pack+Encode: 1.987s, Decode+Unpack: 1.561s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4632 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample152-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample152-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample16-layer4-item1.zst (27/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample16-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 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, 185, 128) -Output shape: (1, 185, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.output: torch.Size([1, 185, 4096]) -> torch.Size([1, 1, 185, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,324B, BPFP=1.0694 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,224B, BPFP=2.8811 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,964B, BPFP=1.9833 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,896B, BPFP=3.1206 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,976B, BPFP=2.3216 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,976B, BPFP=3.2084 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,772B, BPFP=2.3130 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 73,604B, BPFP=3.1083 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,612B, BPFP=2.0106 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,968B, BPFP=3.1659 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 223,732B, BPFP=2.3620 -⌛️ [2/4] FRONTEND: Frontend time: 2.024s (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, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.562s - -[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, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02889625 3.92436062 - layer.0.v_cache 0.00000026 0.00005827 - layer.1.k_cache 0.00305597 0.43814706 - layer.1.v_cache 0.00000081 0.00019738 - layer.2.k_cache 0.00119284 0.16205062 - layer.2.v_cache 0.00000119 0.00029047 - layer.3.k_cache 0.00128722 0.17956489 - layer.3.v_cache 0.00000243 0.00048126 - layer.4.k_cache 0.00400834 0.42683130 - layer.4.v_cache 0.00000313 0.00074212 - layer.4.output 0.00018370 0.02992378 - ------------------------------------------------------------------------------------- - TOTAL 0.00279880 0.37517279 - (elements=2,652,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2652160 -Total Bytes 820048 -BPFP 2.4736 bits/point -EBPFP 4.9472 equivalent bits/point -MSE 0.375173 ----------------------- -------------------------------------------------------- -Time: 3.597s Load: 0.010s, Pack+Encode: 2.024s, Decode+Unpack: 1.562s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3752 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample16-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample16-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample162-layer4-item1.zst (28/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample162-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 139, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 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, 139, 128) -Output shape: (1, 139, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.0.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.1.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.1.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.2.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.2.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.3.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.3.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.4.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.4.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.4.output: torch.Size([1, 139, 4096]) -> torch.Size([1, 1, 139, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 19,408B, BPFP=1.0908 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 63,788B, BPFP=3.5852 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 39,684B, BPFP=2.2304 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 66,840B, BPFP=3.7567 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 47,320B, BPFP=2.6596 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 68,148B, BPFP=3.8303 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 48,444B, BPFP=2.7228 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 66,172B, BPFP=3.7192 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 40,076B, BPFP=2.2525 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 67,708B, BPFP=3.8055 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 186,940B, BPFP=2.6267 -⌛️ [2/4] FRONTEND: Frontend time: 1.977s (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, 139, 128]) - layer.0.v_cache: torch.Size([1, 8, 139, 128]) - layer.1.k_cache: torch.Size([1, 8, 139, 128]) - layer.1.v_cache: torch.Size([1, 8, 139, 128]) - layer.2.k_cache: torch.Size([1, 8, 139, 128]) - layer.2.v_cache: torch.Size([1, 8, 139, 128]) - layer.3.k_cache: torch.Size([1, 8, 139, 128]) - layer.3.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.k_cache: torch.Size([1, 8, 139, 128]) - layer.4.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.output: torch.Size([1, 139, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.543s - -[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, 139, 128]) - layer.0.v_cache: torch.Size([1, 8, 139, 128]) - layer.1.k_cache: torch.Size([1, 8, 139, 128]) - layer.1.v_cache: torch.Size([1, 8, 139, 128]) - layer.2.k_cache: torch.Size([1, 8, 139, 128]) - layer.2.v_cache: torch.Size([1, 8, 139, 128]) - layer.3.k_cache: torch.Size([1, 8, 139, 128]) - layer.3.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.k_cache: torch.Size([1, 8, 139, 128]) - layer.4.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.output: torch.Size([1, 139, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03086843 4.51074658 - layer.0.v_cache 0.00000027 0.00006163 - layer.1.k_cache 0.00318250 0.32996582 - layer.1.v_cache 0.00000093 0.00020346 - layer.2.k_cache 0.00113687 0.16033988 - layer.2.v_cache 0.00000113 0.00027227 - layer.3.k_cache 0.00135365 0.18125972 - layer.3.v_cache 0.00000214 0.00044175 - layer.4.k_cache 0.00345519 0.39884831 - layer.4.v_cache 0.00000310 0.00071061 - layer.4.output 0.00019842 0.03323223 - ------------------------------------------------------------------------------------- - TOTAL 0.00291414 0.40826992 - (elements=1,992,704) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1992704 -Total Bytes 714528 -BPFP 2.8686 bits/point -EBPFP 5.7372 equivalent bits/point -MSE 0.408270 ----------------------- -------------------------------------------------------- -Time: 3.528s Load: 0.008s, Pack+Encode: 1.977s, Decode+Unpack: 1.543s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 139, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4083 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample162-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample162-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample166-layer4-item1.zst (29/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample166-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.014s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,676B, BPFP=1.0300 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,208B, BPFP=3.0527 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,652B, BPFP=2.1190 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,256B, BPFP=3.2366 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,132B, BPFP=2.4588 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,988B, BPFP=3.2698 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,576B, BPFP=2.4789 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,060B, BPFP=3.2277 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,060B, BPFP=2.1375 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,412B, BPFP=3.3345 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 202,804B, BPFP=2.3029 -⌛️ [2/4] FRONTEND: Frontend time: 1.977s (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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.548s - -[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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02678066 4.41726649 - layer.0.v_cache 0.00000028 0.00006376 - layer.1.k_cache 0.00311261 0.41728925 - layer.1.v_cache 0.00000082 0.00019986 - layer.2.k_cache 0.00118278 0.16204297 - layer.2.v_cache 0.00000114 0.00026810 - layer.3.k_cache 0.00135865 0.18962594 - layer.3.v_cache 0.00000220 0.00045108 - layer.4.k_cache 0.00351030 0.47991216 - layer.4.v_cache 0.00000304 0.00069863 - layer.4.output 0.00023867 0.03905828 - ------------------------------------------------------------------------------------- - TOTAL 0.00263622 0.41600367 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 782824 -BPFP 2.5398 bits/point -EBPFP 5.0796 equivalent bits/point -MSE 0.416004 ----------------------- -------------------------------------------------------- -Time: 3.539s Load: 0.014s, Pack+Encode: 1.977s, Decode+Unpack: 1.548s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4160 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample166-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample166-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample175-layer4-item1.zst (30/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample175-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 213, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 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, 213, 128) -Output shape: (1, 213, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.0.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.1.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.1.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.2.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.2.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.3.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.3.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.4.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.4.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.4.output: torch.Size([1, 213, 4096]) -> torch.Size([1, 1, 213, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,324B, BPFP=0.9288 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 86,192B, BPFP=3.1614 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 55,136B, BPFP=2.0223 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 91,036B, BPFP=3.3391 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 65,440B, BPFP=2.4002 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 93,608B, BPFP=3.4334 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 67,240B, BPFP=2.4663 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 90,508B, BPFP=3.3197 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 55,500B, BPFP=2.0357 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 93,244B, BPFP=3.4200 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 236,060B, BPFP=2.1646 -⌛️ [2/4] FRONTEND: Frontend time: 2.387s (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, 213, 128]) - layer.0.v_cache: torch.Size([1, 8, 213, 128]) - layer.1.k_cache: torch.Size([1, 8, 213, 128]) - layer.1.v_cache: torch.Size([1, 8, 213, 128]) - layer.2.k_cache: torch.Size([1, 8, 213, 128]) - layer.2.v_cache: torch.Size([1, 8, 213, 128]) - layer.3.k_cache: torch.Size([1, 8, 213, 128]) - layer.3.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.k_cache: torch.Size([1, 8, 213, 128]) - layer.4.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.output: torch.Size([1, 213, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.993s - -[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, 213, 128]) - layer.0.v_cache: torch.Size([1, 8, 213, 128]) - layer.1.k_cache: torch.Size([1, 8, 213, 128]) - layer.1.v_cache: torch.Size([1, 8, 213, 128]) - layer.2.k_cache: torch.Size([1, 8, 213, 128]) - layer.2.v_cache: torch.Size([1, 8, 213, 128]) - layer.3.k_cache: torch.Size([1, 8, 213, 128]) - layer.3.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.k_cache: torch.Size([1, 8, 213, 128]) - layer.4.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.output: torch.Size([1, 213, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02775377 4.23801447 - layer.0.v_cache 0.00000028 0.00006128 - layer.1.k_cache 0.00299314 0.34277659 - layer.1.v_cache 0.00000076 0.00017650 - layer.2.k_cache 0.00116592 0.17333397 - layer.2.v_cache 0.00000113 0.00024407 - layer.3.k_cache 0.00131338 0.17492264 - layer.3.v_cache 0.00000208 0.00040652 - layer.4.k_cache 0.00376460 0.48557991 - layer.4.v_cache 0.00000302 0.00064808 - layer.4.output 0.00016925 0.03044865 - ------------------------------------------------------------------------------------- - TOTAL 0.00269108 0.39556847 - (elements=3,053,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3053568 -Total Bytes 959288 -BPFP 2.5132 bits/point -EBPFP 5.0265 equivalent bits/point -MSE 0.395568 ----------------------- -------------------------------------------------------- -Time: 4.391s Load: 0.011s, Pack+Encode: 2.387s, Decode+Unpack: 1.993s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 213, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3956 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample175-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample175-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample18-layer4-item1.zst (31/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample18-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 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, 179, 128) -Output shape: (1, 179, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.output: torch.Size([1, 179, 4096]) -> torch.Size([1, 1, 179, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,568B, BPFP=1.1159 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,696B, BPFP=2.9983 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,408B, BPFP=2.0255 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,848B, BPFP=3.2231 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,828B, BPFP=2.3930 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,720B, BPFP=3.3048 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,592B, BPFP=2.3827 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 75,168B, BPFP=3.2807 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,416B, BPFP=2.0695 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,388B, BPFP=3.2903 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 222,644B, BPFP=2.4293 -⌛️ [2/4] FRONTEND: Frontend time: 1.940s (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, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.590s - -[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, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02728191 4.74418018 - layer.0.v_cache 0.00000029 0.00006046 - layer.1.k_cache 0.00306154 0.38688834 - layer.1.v_cache 0.00000084 0.00020284 - layer.2.k_cache 0.00116357 0.15587990 - layer.2.v_cache 0.00000127 0.00027375 - layer.3.k_cache 0.00133958 0.17150102 - layer.3.v_cache 0.00000230 0.00045761 - layer.4.k_cache 0.00349468 0.40215894 - layer.4.v_cache 0.00000328 0.00075484 - layer.4.output 0.00017175 0.02749563 - ------------------------------------------------------------------------------------- - TOTAL 0.00264545 0.42659574 - (elements=2,566,144) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2566144 -Total Bytes 820276 -BPFP 2.5572 bits/point -EBPFP 5.1145 equivalent bits/point -MSE 0.426596 ----------------------- -------------------------------------------------------- -Time: 3.539s Load: 0.009s, Pack+Encode: 1.940s, Decode+Unpack: 1.590s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4266 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample18-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample18-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample19-layer4-item1.zst (32/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample19-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 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, 180, 128) -Output shape: (1, 180, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.output: torch.Size([1, 180, 4096]) -> torch.Size([1, 1, 180, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,620B, BPFP=1.0686 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,728B, BPFP=2.9396 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,528B, BPFP=2.0194 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,656B, BPFP=3.1969 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,688B, BPFP=2.3736 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,056B, BPFP=3.2576 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,752B, BPFP=2.3764 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,552B, BPFP=3.1056 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,892B, BPFP=2.0352 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,132B, BPFP=3.1741 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 203,484B, BPFP=2.2079 -⌛️ [2/4] FRONTEND: Frontend time: 1.976s (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, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.545s - -[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, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03239844 4.37447951 - layer.0.v_cache 0.00000026 0.00005836 - layer.1.k_cache 0.00307858 0.40195380 - layer.1.v_cache 0.00000079 0.00019316 - layer.2.k_cache 0.00118466 0.15967312 - layer.2.v_cache 0.00000116 0.00027951 - layer.3.k_cache 0.00130112 0.17462216 - layer.3.v_cache 0.00000222 0.00045201 - layer.4.k_cache 0.00350099 0.43505881 - layer.4.v_cache 0.00000308 0.00070216 - layer.4.output 0.00016632 0.02606569 - ------------------------------------------------------------------------------------- - TOTAL 0.00300976 0.40369538 - (elements=2,580,480) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2580480 -Total Bytes 792088 -BPFP 2.4556 bits/point -EBPFP 4.9113 equivalent bits/point -MSE 0.403695 ----------------------- -------------------------------------------------------- -Time: 3.531s Load: 0.010s, Pack+Encode: 1.976s, Decode+Unpack: 1.545s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4037 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample19-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample19-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample190-layer4-item1.zst (33/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample190-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 145, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 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, 145, 128) -Output shape: (1, 145, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.0.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.1.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.1.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.2.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.2.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.3.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.3.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.4.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.4.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.4.output: torch.Size([1, 145, 4096]) -> torch.Size([1, 1, 145, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 19,644B, BPFP=1.0584 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 63,848B, BPFP=3.4401 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 40,680B, BPFP=2.1918 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 67,664B, BPFP=3.6457 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 48,580B, BPFP=2.6175 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 68,444B, BPFP=3.6877 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 49,208B, BPFP=2.6513 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 66,540B, BPFP=3.5851 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 40,860B, BPFP=2.2015 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 67,948B, BPFP=3.6610 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 173,892B, BPFP=2.3423 -⌛️ [2/4] FRONTEND: Frontend time: 1.971s (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, 145, 128]) - layer.0.v_cache: torch.Size([1, 8, 145, 128]) - layer.1.k_cache: torch.Size([1, 8, 145, 128]) - layer.1.v_cache: torch.Size([1, 8, 145, 128]) - layer.2.k_cache: torch.Size([1, 8, 145, 128]) - layer.2.v_cache: torch.Size([1, 8, 145, 128]) - layer.3.k_cache: torch.Size([1, 8, 145, 128]) - layer.3.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.k_cache: torch.Size([1, 8, 145, 128]) - layer.4.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.output: torch.Size([1, 145, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.561s - -[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, 145, 128]) - layer.0.v_cache: torch.Size([1, 8, 145, 128]) - layer.1.k_cache: torch.Size([1, 8, 145, 128]) - layer.1.v_cache: torch.Size([1, 8, 145, 128]) - layer.2.k_cache: torch.Size([1, 8, 145, 128]) - layer.2.v_cache: torch.Size([1, 8, 145, 128]) - layer.3.k_cache: torch.Size([1, 8, 145, 128]) - layer.3.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.k_cache: torch.Size([1, 8, 145, 128]) - layer.4.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.output: torch.Size([1, 145, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02766713 4.60710112 - layer.0.v_cache 0.00000028 0.00006118 - layer.1.k_cache 0.00309559 0.32853362 - layer.1.v_cache 0.00000087 0.00020119 - layer.2.k_cache 0.00119714 0.16308725 - layer.2.v_cache 0.00000118 0.00028476 - layer.3.k_cache 0.00133867 0.18011981 - layer.3.v_cache 0.00000208 0.00043643 - layer.4.k_cache 0.00347171 0.40404837 - layer.4.v_cache 0.00000305 0.00068860 - layer.4.output 0.00016146 0.03847946 - ------------------------------------------------------------------------------------- - TOTAL 0.00267311 0.41703430 - (elements=2,078,720) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2078720 -Total Bytes 707308 -BPFP 2.7221 bits/point -EBPFP 5.4442 equivalent bits/point -MSE 0.417034 ----------------------- -------------------------------------------------------- -Time: 3.541s Load: 0.009s, Pack+Encode: 1.971s, Decode+Unpack: 1.561s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 145, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4170 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample190-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample190-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample20-layer4-item1.zst (34/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample20-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 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, 179, 128) -Output shape: (1, 179, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.output: torch.Size([1, 179, 4096]) -> torch.Size([1, 1, 179, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,076B, BPFP=1.0072 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,112B, BPFP=2.9291 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 47,240B, BPFP=2.0618 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 75,764B, BPFP=3.3067 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,580B, BPFP=2.3822 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,888B, BPFP=3.1812 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,988B, BPFP=2.4000 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,204B, BPFP=3.1077 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,100B, BPFP=2.0557 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,672B, BPFP=3.2154 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 205,868B, BPFP=2.2463 -⌛️ [2/4] FRONTEND: Frontend time: 1.949s (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, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.580s - -[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, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02963084 4.11035600 - layer.0.v_cache 0.00000026 0.00005838 - layer.1.k_cache 0.00308074 0.40746520 - layer.1.v_cache 0.00000080 0.00018857 - layer.2.k_cache 0.00119595 0.16332372 - layer.2.v_cache 0.00000115 0.00027406 - layer.3.k_cache 0.00135258 0.18090712 - layer.3.v_cache 0.00000217 0.00042724 - layer.4.k_cache 0.00351189 0.45621981 - layer.4.v_cache 0.00000298 0.00067525 - layer.4.output 0.00021521 0.03303011 - ------------------------------------------------------------------------------------- - TOTAL 0.00283144 0.38942970 - (elements=2,566,144) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2566144 -Total Bytes 793492 -BPFP 2.4737 bits/point -EBPFP 4.9475 equivalent bits/point -MSE 0.389430 ----------------------- -------------------------------------------------------- -Time: 3.540s Load: 0.010s, Pack+Encode: 1.949s, Decode+Unpack: 1.580s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3894 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample20-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample20-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample21-layer4-item1.zst (35/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample21-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: 0.010s - ------------------------------------------------------------- -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: 21,612B, BPFP=1.0754 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 65,600B, BPFP=3.2643 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,892B, BPFP=2.0846 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,080B, BPFP=3.5370 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,912B, BPFP=2.5334 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,156B, BPFP=3.5408 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,664B, BPFP=2.5211 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 69,620B, BPFP=3.4644 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 43,136B, BPFP=2.1465 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,840B, BPFP=3.6246 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 193,852B, BPFP=2.4116 -⌛️ [2/4] FRONTEND: Frontend time: 1.965s (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: 1.572s - -[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 0.02651285 4.02187663 - layer.0.v_cache 0.00000026 0.00005846 - layer.1.k_cache 0.00308002 0.33492196 - layer.1.v_cache 0.00000091 0.00021004 - layer.2.k_cache 0.00121857 0.16703668 - layer.2.v_cache 0.00000134 0.00029202 - layer.3.k_cache 0.00131702 0.17944171 - layer.3.v_cache 0.00000230 0.00046574 - layer.4.k_cache 0.00350090 0.40434732 - layer.4.v_cache 0.00000345 0.00077227 - layer.4.output 0.00019599 0.03697276 - ------------------------------------------------------------------------------------- - TOTAL 0.00260154 0.37552242 - (elements=2,250,752) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2250752 -Total Bytes 752364 -BPFP 2.6742 bits/point -EBPFP 5.3484 equivalent bits/point -MSE 0.375522 ----------------------- -------------------------------------------------------- -Time: 3.547s Load: 0.010s, Pack+Encode: 1.965s, Decode+Unpack: 1.572s ----------------------- -------------------------------------------------------- -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 0.3755 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample21-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample21-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample22-layer4-item1.zst (36/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample22-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 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, 175, 128) -Output shape: (1, 175, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.output: torch.Size([1, 175, 4096]) -> torch.Size([1, 1, 175, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,308B, BPFP=1.0852 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,200B, BPFP=3.0446 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,692B, BPFP=2.0845 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,236B, BPFP=3.3141 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,752B, BPFP=2.4443 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,016B, BPFP=3.3489 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,508B, BPFP=2.4334 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,196B, BPFP=3.2230 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,900B, BPFP=2.0938 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,448B, BPFP=3.3682 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 210,308B, BPFP=2.3472 -⌛️ [2/4] FRONTEND: Frontend time: 1.953s (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, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.600s - -[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, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02748643 4.33738316 - layer.0.v_cache 0.00000027 0.00005871 - layer.1.k_cache 0.00298274 0.39311637 - layer.1.v_cache 0.00000083 0.00020303 - layer.2.k_cache 0.00117838 0.15475031 - layer.2.v_cache 0.00000115 0.00028081 - layer.3.k_cache 0.00127749 0.17411883 - layer.3.v_cache 0.00000229 0.00047021 - layer.4.k_cache 0.00388258 0.43428977 - layer.4.v_cache 0.00000318 0.00072736 - layer.4.output 0.00017171 0.03032376 - ------------------------------------------------------------------------------------- - TOTAL 0.00267873 0.40119240 - (elements=2,508,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2508800 -Total Bytes 802564 -BPFP 2.5592 bits/point -EBPFP 5.1184 equivalent bits/point -MSE 0.401192 ----------------------- -------------------------------------------------------- -Time: 3.562s Load: 0.010s, Pack+Encode: 1.953s, Decode+Unpack: 1.600s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4012 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample22-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample22-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample23-layer4-item1.zst (37/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample23-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 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, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,624B, BPFP=1.0730 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 69,232B, BPFP=3.1446 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,412B, BPFP=2.1081 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,284B, BPFP=3.3287 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,456B, BPFP=2.4735 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,200B, BPFP=3.4157 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,500B, BPFP=2.4755 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,604B, BPFP=3.2978 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,148B, BPFP=2.1415 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,824B, BPFP=3.4440 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 214,608B, BPFP=2.4370 -⌛️ [2/4] FRONTEND: Frontend time: 1.928s (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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.430s - -[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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02961682 4.59892166 - layer.0.v_cache 0.00000027 0.00006376 - layer.1.k_cache 0.00302783 0.39875017 - layer.1.v_cache 0.00000102 0.00021281 - layer.2.k_cache 0.00123476 0.16160340 - layer.2.v_cache 0.00000123 0.00029115 - layer.3.k_cache 0.00126690 0.17120558 - layer.3.v_cache 0.00000235 0.00046869 - layer.4.k_cache 0.00342846 0.38299569 - layer.4.v_cache 0.00000335 0.00077107 - layer.4.output 0.00019350 0.02934799 - ------------------------------------------------------------------------------------- - TOTAL 0.00281121 0.41661971 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 806892 -BPFP 2.6179 bits/point -EBPFP 5.2358 equivalent bits/point -MSE 0.416620 ----------------------- -------------------------------------------------------- -Time: 3.367s Load: 0.009s, Pack+Encode: 1.928s, Decode+Unpack: 1.430s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4166 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample23-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample23-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample24-layer4-item1.zst (38/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample24-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: 24,224B, BPFP=1.0632 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,588B, BPFP=2.9665 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,188B, BPFP=2.0272 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,400B, BPFP=3.1777 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,092B, BPFP=2.3741 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,392B, BPFP=3.2651 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,568B, BPFP=2.3950 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,128B, BPFP=3.1657 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,132B, BPFP=2.0686 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,076B, BPFP=3.2073 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 200,080B, BPFP=2.1954 -⌛️ [2/4] FRONTEND: Frontend time: 1.930s (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: 1.579s - -[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, 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 0.02741563 4.63204750 - layer.0.v_cache 0.00000026 0.00006368 - layer.1.k_cache 0.00307030 0.34093379 - layer.1.v_cache 0.00000077 0.00019795 - layer.2.k_cache 0.00116729 0.16765474 - layer.2.v_cache 0.00000113 0.00027339 - layer.3.k_cache 0.00134025 0.18180003 - layer.3.v_cache 0.00000215 0.00045216 - layer.4.k_cache 0.00345504 0.43933478 - layer.4.v_cache 0.00000307 0.00073126 - layer.4.output 0.00018601 0.02697120 - ------------------------------------------------------------------------------------- - TOTAL 0.00265714 0.41938386 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 785868 -BPFP 2.4637 bits/point -EBPFP 4.9274 equivalent bits/point -MSE 0.419384 ----------------------- -------------------------------------------------------- -Time: 3.519s Load: 0.010s, Pack+Encode: 1.930s, Decode+Unpack: 1.579s ----------------------- -------------------------------------------------------- -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 0.4194 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample24-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample24-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample25-layer4-item1.zst (39/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample25-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 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, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,296B, BPFP=1.0456 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,180B, BPFP=2.8183 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,708B, BPFP=1.9307 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 77,176B, BPFP=3.1901 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 55,048B, BPFP=2.2755 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 78,084B, BPFP=3.2277 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 55,256B, BPFP=2.2841 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 74,004B, BPFP=3.0590 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,148B, BPFP=1.9489 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,876B, BPFP=3.1364 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 216,016B, BPFP=2.2323 -⌛️ [2/4] FRONTEND: Frontend time: 1.983s (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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.580s - -[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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02569163 4.26733624 - layer.0.v_cache 0.00000027 0.00005916 - layer.1.k_cache 0.00297363 0.43162113 - layer.1.v_cache 0.00000081 0.00019742 - layer.2.k_cache 0.00119805 0.16021791 - layer.2.v_cache 0.00000116 0.00027195 - layer.3.k_cache 0.00128923 0.17033554 - layer.3.v_cache 0.00000232 0.00046583 - layer.4.k_cache 0.00350442 0.41143823 - layer.4.v_cache 0.00000328 0.00072086 - layer.4.output 0.00017922 0.03525941 - ------------------------------------------------------------------------------------- - TOTAL 0.00252727 0.39883585 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 818792 -BPFP 2.4175 bits/point -EBPFP 4.8351 equivalent bits/point -MSE 0.398836 ----------------------- -------------------------------------------------------- -Time: 3.575s Load: 0.011s, Pack+Encode: 1.983s, Decode+Unpack: 1.580s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3988 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample25-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample25-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample26-layer4-item1.zst (40/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample26-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: 0.011s - ------------------------------------------------------------- -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: 25,084B, BPFP=1.1009 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 69,024B, BPFP=3.0295 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,636B, BPFP=2.0469 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 75,024B, BPFP=3.2928 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,520B, BPFP=2.3929 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,004B, BPFP=3.2920 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,636B, BPFP=2.3980 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,456B, BPFP=3.1362 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,036B, BPFP=2.0644 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,384B, BPFP=3.2209 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 205,552B, BPFP=2.2554 -⌛️ [2/4] FRONTEND: Frontend time: 1.984s (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: 1.566s - -[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 0.02863186 4.60153438 - layer.0.v_cache 0.00000027 0.00005955 - layer.1.k_cache 0.00302134 0.38375914 - layer.1.v_cache 0.00000082 0.00019871 - layer.2.k_cache 0.00115360 0.15999794 - layer.2.v_cache 0.00000119 0.00026832 - layer.3.k_cache 0.00131482 0.17976189 - layer.3.v_cache 0.00000222 0.00044730 - layer.4.k_cache 0.00345682 0.42380755 - layer.4.v_cache 0.00000324 0.00072142 - layer.4.output 0.00020381 0.02984416 - ------------------------------------------------------------------------------------- - TOTAL 0.00274296 0.41928092 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 797356 -BPFP 2.4997 bits/point -EBPFP 4.9995 equivalent bits/point -MSE 0.419281 ----------------------- -------------------------------------------------------- -Time: 3.561s Load: 0.011s, Pack+Encode: 1.984s, Decode+Unpack: 1.566s ----------------------- -------------------------------------------------------- -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 0.4193 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample26-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample26-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample27-layer4-item1.zst (41/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample27-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 182, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 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, 182, 128) -Output shape: (1, 182, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.0.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.1.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.1.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.2.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.2.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.3.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.3.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.4.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.4.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.4.output: torch.Size([1, 182, 4096]) -> torch.Size([1, 1, 182, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,204B, BPFP=1.0819 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,292B, BPFP=2.9315 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,628B, BPFP=2.0015 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,492B, BPFP=3.1976 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,696B, BPFP=2.3479 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 77,268B, BPFP=3.3168 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,768B, BPFP=2.3510 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,864B, BPFP=3.1277 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,984B, BPFP=2.0168 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,484B, BPFP=3.2402 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 208,628B, BPFP=2.2389 -⌛️ [2/4] FRONTEND: Frontend time: 1.980s (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, 182, 128]) - layer.0.v_cache: torch.Size([1, 8, 182, 128]) - layer.1.k_cache: torch.Size([1, 8, 182, 128]) - layer.1.v_cache: torch.Size([1, 8, 182, 128]) - layer.2.k_cache: torch.Size([1, 8, 182, 128]) - layer.2.v_cache: torch.Size([1, 8, 182, 128]) - layer.3.k_cache: torch.Size([1, 8, 182, 128]) - layer.3.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.k_cache: torch.Size([1, 8, 182, 128]) - layer.4.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.output: torch.Size([1, 182, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.558s - -[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, 182, 128]) - layer.0.v_cache: torch.Size([1, 8, 182, 128]) - layer.1.k_cache: torch.Size([1, 8, 182, 128]) - layer.1.v_cache: torch.Size([1, 8, 182, 128]) - layer.2.k_cache: torch.Size([1, 8, 182, 128]) - layer.2.v_cache: torch.Size([1, 8, 182, 128]) - layer.3.k_cache: torch.Size([1, 8, 182, 128]) - layer.3.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.k_cache: torch.Size([1, 8, 182, 128]) - layer.4.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.output: torch.Size([1, 182, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02794281 4.76143705 - layer.0.v_cache 0.00000027 0.00006089 - layer.1.k_cache 0.00307012 0.41659906 - layer.1.v_cache 0.00000079 0.00019753 - layer.2.k_cache 0.00119021 0.15834819 - layer.2.v_cache 0.00000113 0.00027443 - layer.3.k_cache 0.00130892 0.17165406 - layer.3.v_cache 0.00000221 0.00045770 - layer.4.k_cache 0.00347814 0.44593241 - layer.4.v_cache 0.00000308 0.00071782 - layer.4.output 0.00018397 0.02975210 - ------------------------------------------------------------------------------------- - TOTAL 0.00269526 0.43390625 - (elements=2,609,152) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2609152 -Total Bytes 805308 -BPFP 2.4692 bits/point -EBPFP 4.9384 equivalent bits/point -MSE 0.433906 ----------------------- -------------------------------------------------------- -Time: 3.549s Load: 0.011s, Pack+Encode: 1.980s, Decode+Unpack: 1.558s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 182, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4339 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample27-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample27-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample3-layer4-item1.zst (42/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample3-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 271, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.014s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 271, 128) -Output shape: (1, 271, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.0.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.1.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.1.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.2.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.2.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.3.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.3.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.4.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.4.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.4.output: torch.Size([1, 271, 4096]) -> torch.Size([1, 1, 271, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 35,672B, BPFP=1.0284 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 106,948B, BPFP=3.0831 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 68,792B, BPFP=1.9832 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 114,776B, BPFP=3.3088 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 81,956B, BPFP=2.3627 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 115,768B, BPFP=3.3374 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 83,396B, BPFP=2.4042 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 113,192B, BPFP=3.2631 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 68,164B, BPFP=1.9651 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 120,560B, BPFP=3.4756 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 292,688B, BPFP=2.1094 -⌛️ [2/4] FRONTEND: Frontend time: 2.389s (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, 271, 128]) - layer.0.v_cache: torch.Size([1, 8, 271, 128]) - layer.1.k_cache: torch.Size([1, 8, 271, 128]) - layer.1.v_cache: torch.Size([1, 8, 271, 128]) - layer.2.k_cache: torch.Size([1, 8, 271, 128]) - layer.2.v_cache: torch.Size([1, 8, 271, 128]) - layer.3.k_cache: torch.Size([1, 8, 271, 128]) - layer.3.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.k_cache: torch.Size([1, 8, 271, 128]) - layer.4.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.output: torch.Size([1, 271, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.017s - -[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, 271, 128]) - layer.0.v_cache: torch.Size([1, 8, 271, 128]) - layer.1.k_cache: torch.Size([1, 8, 271, 128]) - layer.1.v_cache: torch.Size([1, 8, 271, 128]) - layer.2.k_cache: torch.Size([1, 8, 271, 128]) - layer.2.v_cache: torch.Size([1, 8, 271, 128]) - layer.3.k_cache: torch.Size([1, 8, 271, 128]) - layer.3.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.k_cache: torch.Size([1, 8, 271, 128]) - layer.4.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.output: torch.Size([1, 271, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02401601 4.11218667 - layer.0.v_cache 0.00000027 0.00005728 - layer.1.k_cache 0.00298786 0.34597201 - layer.1.v_cache 0.00000080 0.00017885 - layer.2.k_cache 0.00117700 0.15591334 - layer.2.v_cache 0.00000117 0.00025214 - layer.3.k_cache 0.00127449 0.16671942 - layer.3.v_cache 0.00000214 0.00041406 - layer.4.k_cache 0.00368668 0.39026000 - layer.4.v_cache 0.00000306 0.00066505 - layer.4.output 0.00014564 0.02860669 - ------------------------------------------------------------------------------------- - TOTAL 0.00240943 0.37764611 - (elements=3,885,056) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3885056 -Total Bytes 1201912 -BPFP 2.4749 bits/point -EBPFP 4.9499 equivalent bits/point -MSE 0.377646 ----------------------- -------------------------------------------------------- -Time: 4.420s Load: 0.014s, Pack+Encode: 2.389s, Decode+Unpack: 2.017s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 271, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3776 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample3-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample3-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample30-layer4-item1.zst (43/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample30-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 184, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 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, 184, 128) -Output shape: (1, 184, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.0.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.1.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.1.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.2.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.2.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.3.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.3.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.4.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.4.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.4.output: torch.Size([1, 184, 4096]) -> torch.Size([1, 1, 184, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 26,464B, BPFP=1.1236 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,972B, BPFP=2.9285 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,496B, BPFP=1.9742 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,984B, BPFP=3.1838 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,320B, BPFP=2.3064 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,172B, BPFP=3.1917 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 55,644B, BPFP=2.3626 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,072B, BPFP=3.0601 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,520B, BPFP=2.0177 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,524B, BPFP=3.1642 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 242,392B, BPFP=2.5729 -⌛️ [2/4] FRONTEND: Frontend time: 1.994s (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, 184, 128]) - layer.0.v_cache: torch.Size([1, 8, 184, 128]) - layer.1.k_cache: torch.Size([1, 8, 184, 128]) - layer.1.v_cache: torch.Size([1, 8, 184, 128]) - layer.2.k_cache: torch.Size([1, 8, 184, 128]) - layer.2.v_cache: torch.Size([1, 8, 184, 128]) - layer.3.k_cache: torch.Size([1, 8, 184, 128]) - layer.3.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.k_cache: torch.Size([1, 8, 184, 128]) - layer.4.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.output: torch.Size([1, 184, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.544s - -[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, 184, 128]) - layer.0.v_cache: torch.Size([1, 8, 184, 128]) - layer.1.k_cache: torch.Size([1, 8, 184, 128]) - layer.1.v_cache: torch.Size([1, 8, 184, 128]) - layer.2.k_cache: torch.Size([1, 8, 184, 128]) - layer.2.v_cache: torch.Size([1, 8, 184, 128]) - layer.3.k_cache: torch.Size([1, 8, 184, 128]) - layer.3.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.k_cache: torch.Size([1, 8, 184, 128]) - layer.4.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.output: torch.Size([1, 184, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02872292 5.47600124 - layer.0.v_cache 0.00000026 0.00005712 - layer.1.k_cache 0.00314513 0.40099281 - layer.1.v_cache 0.00000084 0.00018786 - layer.2.k_cache 0.00123314 0.17301468 - layer.2.v_cache 0.00000120 0.00027551 - layer.3.k_cache 0.00131532 0.17860740 - layer.3.v_cache 0.00000224 0.00043151 - layer.4.k_cache 0.00360741 0.45866552 - layer.4.v_cache 0.00000320 0.00069583 - layer.4.output 0.00023634 0.03200768 - ------------------------------------------------------------------------------------- - TOTAL 0.00278407 0.48692573 - (elements=2,637,824) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2637824 -Total Bytes 838560 -BPFP 2.5432 bits/point -EBPFP 5.0864 equivalent bits/point -MSE 0.486926 ----------------------- -------------------------------------------------------- -Time: 3.548s Load: 0.011s, Pack+Encode: 1.994s, Decode+Unpack: 1.544s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 184, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4869 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample30-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample30-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample31-layer4-item1.zst (44/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample31-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 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, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,192B, BPFP=1.1118 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,744B, BPFP=3.1592 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,860B, BPFP=2.1075 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,788B, BPFP=3.3450 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,244B, BPFP=2.4928 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,668B, BPFP=3.4314 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,540B, BPFP=2.5064 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,444B, BPFP=3.3292 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,012B, BPFP=2.1605 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,404B, BPFP=3.4653 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 210,232B, BPFP=2.4153 -⌛️ [2/4] FRONTEND: Frontend time: 1.946s (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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.565s - -[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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02895958 4.52451423 - layer.0.v_cache 0.00000027 0.00005741 - layer.1.k_cache 0.00314466 0.39742427 - layer.1.v_cache 0.00000087 0.00020717 - layer.2.k_cache 0.00118284 0.16116093 - layer.2.v_cache 0.00000133 0.00028097 - layer.3.k_cache 0.00130601 0.17830352 - layer.3.v_cache 0.00000238 0.00047261 - layer.4.k_cache 0.00330381 0.43839084 - layer.4.v_cache 0.00000327 0.00074663 - layer.4.output 0.00019227 0.03157586 - ------------------------------------------------------------------------------------- - TOTAL 0.00276244 0.41627586 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 800128 -BPFP 2.6265 bits/point -EBPFP 5.2529 equivalent bits/point -MSE 0.416276 ----------------------- -------------------------------------------------------- -Time: 3.521s Load: 0.010s, Pack+Encode: 1.946s, Decode+Unpack: 1.565s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4163 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample31-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample31-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample32-layer4-item1.zst (45/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample32-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 206, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 206, 128) -Output shape: (1, 206, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.0.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.1.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.1.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.2.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.2.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.3.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.3.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.4.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.4.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.4.output: torch.Size([1, 206, 4096]) -> torch.Size([1, 1, 206, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 28,280B, BPFP=1.0725 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 85,864B, BPFP=3.2564 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 54,232B, BPFP=2.0567 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 93,492B, BPFP=3.5457 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 63,380B, BPFP=2.4037 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 92,796B, BPFP=3.5193 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 66,992B, BPFP=2.5407 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 91,652B, BPFP=3.4759 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 54,276B, BPFP=2.0584 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 92,752B, BPFP=3.5176 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 285,172B, BPFP=2.7038 -⌛️ [2/4] FRONTEND: Frontend time: 2.396s (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, 206, 128]) - layer.0.v_cache: torch.Size([1, 8, 206, 128]) - layer.1.k_cache: torch.Size([1, 8, 206, 128]) - layer.1.v_cache: torch.Size([1, 8, 206, 128]) - layer.2.k_cache: torch.Size([1, 8, 206, 128]) - layer.2.v_cache: torch.Size([1, 8, 206, 128]) - layer.3.k_cache: torch.Size([1, 8, 206, 128]) - layer.3.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.k_cache: torch.Size([1, 8, 206, 128]) - layer.4.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.output: torch.Size([1, 206, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.938s - -[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, 206, 128]) - layer.0.v_cache: torch.Size([1, 8, 206, 128]) - layer.1.k_cache: torch.Size([1, 8, 206, 128]) - layer.1.v_cache: torch.Size([1, 8, 206, 128]) - layer.2.k_cache: torch.Size([1, 8, 206, 128]) - layer.2.v_cache: torch.Size([1, 8, 206, 128]) - layer.3.k_cache: torch.Size([1, 8, 206, 128]) - layer.3.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.k_cache: torch.Size([1, 8, 206, 128]) - layer.4.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.output: torch.Size([1, 206, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02717657 4.12863663 - layer.0.v_cache 0.00000026 0.00005993 - layer.1.k_cache 0.00302264 0.32000281 - layer.1.v_cache 0.00000082 0.00019751 - layer.2.k_cache 0.00113561 0.15882279 - layer.2.v_cache 0.00000128 0.00026535 - layer.3.k_cache 0.00127405 0.17258994 - layer.3.v_cache 0.00000234 0.00045857 - layer.4.k_cache 0.00358677 0.43556828 - layer.4.v_cache 0.00000306 0.00067555 - layer.4.output 0.00018722 0.03471859 - ------------------------------------------------------------------------------------- - TOTAL 0.00263945 0.38258226 - (elements=2,953,216) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2953216 -Total Bytes 1008888 -BPFP 2.7330 bits/point -EBPFP 5.4660 equivalent bits/point -MSE 0.382582 ----------------------- -------------------------------------------------------- -Time: 4.346s Load: 0.012s, Pack+Encode: 2.396s, Decode+Unpack: 1.938s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 206, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3826 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample32-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample32-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample33-layer4-item1.zst (46/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample33-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 191, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 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, 191, 128) -Output shape: (1, 191, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.0.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.1.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.1.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.2.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.2.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.3.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.3.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.4.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.4.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.4.output: torch.Size([1, 191, 4096]) -> torch.Size([1, 1, 191, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 30,408B, BPFP=1.2438 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,420B, BPFP=2.7986 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,224B, BPFP=1.8498 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 81,252B, BPFP=3.3235 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,864B, BPFP=2.2032 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 81,336B, BPFP=3.3269 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,740B, BPFP=2.1981 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 73,124B, BPFP=2.9910 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,668B, BPFP=1.8680 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 83,412B, BPFP=3.4118 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 212,988B, BPFP=2.1780 -⌛️ [2/4] FRONTEND: Frontend time: 1.965s (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, 191, 128]) - layer.0.v_cache: torch.Size([1, 8, 191, 128]) - layer.1.k_cache: torch.Size([1, 8, 191, 128]) - layer.1.v_cache: torch.Size([1, 8, 191, 128]) - layer.2.k_cache: torch.Size([1, 8, 191, 128]) - layer.2.v_cache: torch.Size([1, 8, 191, 128]) - layer.3.k_cache: torch.Size([1, 8, 191, 128]) - layer.3.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.k_cache: torch.Size([1, 8, 191, 128]) - layer.4.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.output: torch.Size([1, 191, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.560s - -[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, 191, 128]) - layer.0.v_cache: torch.Size([1, 8, 191, 128]) - layer.1.k_cache: torch.Size([1, 8, 191, 128]) - layer.1.v_cache: torch.Size([1, 8, 191, 128]) - layer.2.k_cache: torch.Size([1, 8, 191, 128]) - layer.2.v_cache: torch.Size([1, 8, 191, 128]) - layer.3.k_cache: torch.Size([1, 8, 191, 128]) - layer.3.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.k_cache: torch.Size([1, 8, 191, 128]) - layer.4.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.output: torch.Size([1, 191, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02702075 6.18989667 - layer.0.v_cache 0.00000026 0.00005779 - layer.1.k_cache 0.00310674 0.39726904 - layer.1.v_cache 0.00000076 0.00017948 - layer.2.k_cache 0.00119798 0.15571594 - layer.2.v_cache 0.00000111 0.00024889 - layer.3.k_cache 0.00132973 0.16681726 - layer.3.v_cache 0.00000216 0.00042245 - layer.4.k_cache 0.00342601 0.41293942 - layer.4.v_cache 0.00000312 0.00068744 - layer.4.output 0.00019316 0.02592572 - ------------------------------------------------------------------------------------- - TOTAL 0.00263295 0.53056695 - (elements=2,738,176) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2738176 -Total Bytes 829436 -BPFP 2.4233 bits/point -EBPFP 4.8466 equivalent bits/point -MSE 0.530567 ----------------------- -------------------------------------------------------- -Time: 3.536s Load: 0.011s, Pack+Encode: 1.965s, Decode+Unpack: 1.560s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 191, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.5306 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample33-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample33-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample35-layer4-item1.zst (47/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample35-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: 0.010s - ------------------------------------------------------------- -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: 23,836B, BPFP=1.0764 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,872B, BPFP=3.0650 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,376B, BPFP=2.0943 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,560B, BPFP=3.3219 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,084B, BPFP=2.4424 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,960B, BPFP=3.2948 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,400B, BPFP=2.4566 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,428B, BPFP=3.2256 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,000B, BPFP=2.1225 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 78,492B, BPFP=3.5446 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 207,188B, BPFP=2.3391 -⌛️ [2/4] FRONTEND: Frontend time: 1.982s (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: 1.544s - -[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 0.02805312 4.17422715 - layer.0.v_cache 0.00000027 0.00005773 - layer.1.k_cache 0.00341978 0.42161613 - layer.1.v_cache 0.00000082 0.00020690 - layer.2.k_cache 0.00114483 0.16065884 - layer.2.v_cache 0.00000117 0.00028305 - layer.3.k_cache 0.00131871 0.17572448 - layer.3.v_cache 0.00000219 0.00045453 - layer.4.k_cache 0.00345933 0.42829604 - layer.4.v_cache 0.00000316 0.00074628 - layer.4.output 0.00017521 0.03512668 - ------------------------------------------------------------------------------------- - TOTAL 0.00272173 0.39305556 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 797196 -BPFP 2.5715 bits/point -EBPFP 5.1429 equivalent bits/point -MSE 0.393056 ----------------------- -------------------------------------------------------- -Time: 3.536s Load: 0.010s, Pack+Encode: 1.982s, Decode+Unpack: 1.544s ----------------------- -------------------------------------------------------- -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 0.3931 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample35-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample35-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample36-layer4-item1.zst (48/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample36-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 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, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,228B, BPFP=1.0738 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,256B, BPFP=3.1553 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,872B, BPFP=2.0743 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,160B, BPFP=3.3358 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,768B, BPFP=2.4856 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,520B, BPFP=3.3987 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,400B, BPFP=2.5148 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,480B, BPFP=3.3044 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,596B, BPFP=2.1540 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,024B, BPFP=3.4682 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 204,488B, BPFP=2.3633 -⌛️ [2/4] FRONTEND: Frontend time: 1.940s (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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.536s - -[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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02802299 4.09453876 - layer.0.v_cache 0.00000027 0.00006010 - layer.1.k_cache 0.00309164 0.33333515 - layer.1.v_cache 0.00000083 0.00020128 - layer.2.k_cache 0.00116255 0.16475373 - layer.2.v_cache 0.00000115 0.00027347 - layer.3.k_cache 0.00130840 0.18592784 - layer.3.v_cache 0.00000220 0.00045308 - layer.4.k_cache 0.00347752 0.46968038 - layer.4.v_cache 0.00000302 0.00070546 - layer.4.output 0.00021363 0.03176134 - ------------------------------------------------------------------------------------- - TOTAL 0.00270893 0.38406961 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 787792 -BPFP 2.6013 bits/point -EBPFP 5.2026 equivalent bits/point -MSE 0.384070 ----------------------- -------------------------------------------------------- -Time: 3.486s Load: 0.010s, Pack+Encode: 1.940s, Decode+Unpack: 1.536s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3841 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample36-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample36-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample37-layer4-item1.zst (49/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample37-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 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, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 27,308B, BPFP=1.1288 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 69,164B, BPFP=2.8590 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,672B, BPFP=1.9292 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 77,104B, BPFP=3.1872 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,632B, BPFP=2.2583 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 76,744B, BPFP=3.1723 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,912B, BPFP=2.2698 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,560B, BPFP=2.9580 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,156B, BPFP=1.9492 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 78,752B, BPFP=3.2553 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 225,620B, BPFP=2.3316 -⌛️ [2/4] FRONTEND: Frontend time: 1.973s (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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.567s - -[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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02754041 5.43474018 - layer.0.v_cache 0.00000027 0.00005927 - layer.1.k_cache 0.00311113 0.41497548 - layer.1.v_cache 0.00000084 0.00019368 - layer.2.k_cache 0.00124070 0.16575847 - layer.2.v_cache 0.00000144 0.00028022 - layer.3.k_cache 0.00134127 0.17760861 - layer.3.v_cache 0.00000226 0.00044943 - layer.4.k_cache 0.00350879 0.45634045 - layer.4.v_cache 0.00000320 0.00072362 - layer.4.output 0.00022464 0.04201191 - ------------------------------------------------------------------------------------- - TOTAL 0.00268921 0.48708408 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 829624 -BPFP 2.4495 bits/point -EBPFP 4.8990 equivalent bits/point -MSE 0.487084 ----------------------- -------------------------------------------------------- -Time: 3.550s Load: 0.010s, Pack+Encode: 1.973s, Decode+Unpack: 1.567s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4871 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample37-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample37-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample38-layer4-item1.zst (50/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample38-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 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, 177, 128) -Output shape: (1, 177, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.output: torch.Size([1, 177, 4096]) -> torch.Size([1, 1, 177, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,284B, BPFP=1.0719 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,384B, BPFP=2.9742 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,552B, BPFP=2.0547 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,400B, BPFP=3.1956 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,796B, BPFP=2.3745 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,688B, BPFP=3.2083 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,380B, BPFP=2.4002 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,744B, BPFP=3.1225 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,684B, BPFP=2.0606 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,736B, BPFP=3.2546 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 207,128B, BPFP=2.2856 -⌛️ [2/4] FRONTEND: Frontend time: 1.929s (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, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.527s - -[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, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02865209 4.51056529 - layer.0.v_cache 0.00000026 0.00005634 - layer.1.k_cache 0.00315540 0.41702085 - layer.1.v_cache 0.00000078 0.00018760 - layer.2.k_cache 0.00119797 0.17132632 - layer.2.v_cache 0.00000110 0.00026509 - layer.3.k_cache 0.00133161 0.17810747 - layer.3.v_cache 0.00000204 0.00041039 - layer.4.k_cache 0.00358407 0.45079480 - layer.4.v_cache 0.00000307 0.00069499 - layer.4.output 0.00019353 0.02825007 - ------------------------------------------------------------------------------------- - TOTAL 0.00276446 0.41731639 - (elements=2,537,472) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2537472 -Total Bytes 789776 -BPFP 2.4900 bits/point -EBPFP 4.9799 equivalent bits/point -MSE 0.417316 ----------------------- -------------------------------------------------------- -Time: 3.466s Load: 0.010s, Pack+Encode: 1.929s, Decode+Unpack: 1.527s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4173 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample38-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample38-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample39-layer4-item1.zst (51/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample39-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: 23,608B, BPFP=1.0661 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,204B, BPFP=3.0800 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,040B, BPFP=2.0791 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,324B, BPFP=3.3564 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,876B, BPFP=2.4330 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 76,132B, BPFP=3.4380 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,020B, BPFP=2.4395 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,864B, BPFP=3.2905 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,724B, BPFP=2.1100 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,140B, BPFP=3.3029 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 226,124B, BPFP=2.5529 -⌛️ [2/4] FRONTEND: Frontend time: 2.004s (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: 1.531s - -[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 0.02612204 4.09681940 - layer.0.v_cache 0.00000027 0.00005776 - layer.1.k_cache 0.00302657 0.40593340 - layer.1.v_cache 0.00000086 0.00020454 - layer.2.k_cache 0.00117788 0.16110046 - layer.2.v_cache 0.00000116 0.00028300 - layer.3.k_cache 0.00128963 0.17375274 - layer.3.v_cache 0.00000233 0.00045940 - layer.4.k_cache 0.00354125 0.43601469 - layer.4.v_cache 0.00000310 0.00070414 - layer.4.output 0.00019853 0.03289819 - ------------------------------------------------------------------------------------- - TOTAL 0.00256851 0.38620874 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 815056 -BPFP 2.6291 bits/point -EBPFP 5.2582 equivalent bits/point -MSE 0.386209 ----------------------- -------------------------------------------------------- -Time: 3.547s Load: 0.011s, Pack+Encode: 2.004s, Decode+Unpack: 1.531s ----------------------- -------------------------------------------------------- -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 0.3862 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample39-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample39-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample4-layer4-item1.zst (52/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample4-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 243, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 243, 128) -Output shape: (1, 243, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.0.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.1.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.1.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.2.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.2.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.3.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.3.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.4.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.4.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.4.output: torch.Size([1, 243, 4096]) -> torch.Size([1, 1, 243, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 33,896B, BPFP=1.0898 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 90,088B, BPFP=2.8963 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 62,356B, BPFP=2.0048 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 99,148B, BPFP=3.1876 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 72,228B, BPFP=2.3221 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 101,232B, BPFP=3.2546 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 72,004B, BPFP=2.3149 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 97,784B, BPFP=3.1438 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 62,420B, BPFP=2.0068 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 99,672B, BPFP=3.2045 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 285,164B, BPFP=2.2920 -⌛️ [2/4] FRONTEND: Frontend time: 2.398s (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, 243, 128]) - layer.0.v_cache: torch.Size([1, 8, 243, 128]) - layer.1.k_cache: torch.Size([1, 8, 243, 128]) - layer.1.v_cache: torch.Size([1, 8, 243, 128]) - layer.2.k_cache: torch.Size([1, 8, 243, 128]) - layer.2.v_cache: torch.Size([1, 8, 243, 128]) - layer.3.k_cache: torch.Size([1, 8, 243, 128]) - layer.3.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.k_cache: torch.Size([1, 8, 243, 128]) - layer.4.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.output: torch.Size([1, 243, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.944s - -[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, 243, 128]) - layer.0.v_cache: torch.Size([1, 8, 243, 128]) - layer.1.k_cache: torch.Size([1, 8, 243, 128]) - layer.1.v_cache: torch.Size([1, 8, 243, 128]) - layer.2.k_cache: torch.Size([1, 8, 243, 128]) - layer.2.v_cache: torch.Size([1, 8, 243, 128]) - layer.3.k_cache: torch.Size([1, 8, 243, 128]) - layer.3.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.k_cache: torch.Size([1, 8, 243, 128]) - layer.4.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.output: torch.Size([1, 243, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02455453 5.04768227 - layer.0.v_cache 0.00000027 0.00005878 - layer.1.k_cache 0.00295349 0.40673087 - layer.1.v_cache 0.00000082 0.00020560 - layer.2.k_cache 0.00119501 0.15735764 - layer.2.v_cache 0.00000127 0.00029626 - layer.3.k_cache 0.00131209 0.16966976 - layer.3.v_cache 0.00000246 0.00049174 - layer.4.k_cache 0.00353116 0.41952895 - layer.4.v_cache 0.00000321 0.00070941 - layer.4.output 0.00017298 0.02377495 - ------------------------------------------------------------------------------------- - TOTAL 0.00244616 0.44984508 - (elements=3,483,648) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3483648 -Total Bytes 1075992 -BPFP 2.4710 bits/point -EBPFP 4.9419 equivalent bits/point -MSE 0.449845 ----------------------- -------------------------------------------------------- -Time: 4.356s Load: 0.013s, Pack+Encode: 2.398s, Decode+Unpack: 1.944s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 243, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4498 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample4-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample4-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample40-layer4-item1.zst (53/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample40-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 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, 152, 128) -Output shape: (1, 152, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.output: torch.Size([1, 152, 4096]) -> torch.Size([1, 1, 152, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 20,620B, BPFP=1.0598 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 65,132B, BPFP=3.3477 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,344B, BPFP=2.1250 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 69,520B, BPFP=3.5732 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,068B, BPFP=2.5734 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,340B, BPFP=3.6667 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,184B, BPFP=2.5794 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,396B, BPFP=3.5154 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 41,784B, BPFP=2.1476 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 70,596B, BPFP=3.6285 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 195,096B, BPFP=2.5069 -⌛️ [2/4] FRONTEND: Frontend time: 1.985s (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, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.565s - -[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, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02793655 4.25378498 - layer.0.v_cache 0.00000027 0.00006375 - layer.1.k_cache 0.00308216 0.31452335 - layer.1.v_cache 0.00000096 0.00019606 - layer.2.k_cache 0.00119397 0.16330668 - layer.2.v_cache 0.00000115 0.00027289 - layer.3.k_cache 0.00132015 0.17793501 - layer.3.v_cache 0.00000229 0.00046880 - layer.4.k_cache 0.00355994 0.39701540 - layer.4.v_cache 0.00000322 0.00074379 - layer.4.output 0.00016979 0.03541540 - ------------------------------------------------------------------------------------- - TOTAL 0.00269856 0.38928374 - (elements=2,179,072) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2179072 -Total Bytes 744080 -BPFP 2.7317 bits/point -EBPFP 5.4635 equivalent bits/point -MSE 0.389284 ----------------------- -------------------------------------------------------- -Time: 3.558s Load: 0.008s, Pack+Encode: 1.985s, Decode+Unpack: 1.565s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3893 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample40-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample40-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample41-layer4-item1.zst (54/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample41-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 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, 185, 128) -Output shape: (1, 185, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.output: torch.Size([1, 185, 4096]) -> torch.Size([1, 1, 185, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,460B, BPFP=1.0752 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 69,240B, BPFP=2.9240 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,652B, BPFP=1.9701 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 76,696B, BPFP=3.2389 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,720B, BPFP=2.3108 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 76,112B, BPFP=3.2142 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,880B, BPFP=2.3176 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 74,392B, BPFP=3.1416 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,812B, BPFP=1.9769 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,640B, BPFP=3.1520 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 207,856B, BPFP=2.1944 -⌛️ [2/4] FRONTEND: Frontend time: 1.967s (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, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.570s - -[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, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02732194 4.15118441 - layer.0.v_cache 0.00000026 0.00006041 - layer.1.k_cache 0.00316422 0.43563307 - layer.1.v_cache 0.00000081 0.00020175 - layer.2.k_cache 0.00119231 0.15649160 - layer.2.v_cache 0.00000116 0.00028274 - layer.3.k_cache 0.00128139 0.17595693 - layer.3.v_cache 0.00000230 0.00046781 - layer.4.k_cache 0.00354606 0.43026123 - layer.4.v_cache 0.00000329 0.00072848 - layer.4.output 0.00016551 0.03128254 - ------------------------------------------------------------------------------------- - TOTAL 0.00265541 0.39117133 - (elements=2,652,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2652160 -Total Bytes 807460 -BPFP 2.4356 bits/point -EBPFP 4.8713 equivalent bits/point -MSE 0.391171 ----------------------- -------------------------------------------------------- -Time: 3.546s Load: 0.009s, Pack+Encode: 1.967s, Decode+Unpack: 1.570s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3912 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample41-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample41-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample42-layer4-item1.zst (55/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample42-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 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, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,624B, BPFP=1.0459 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,764B, BPFP=3.1326 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,200B, BPFP=2.0895 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,976B, BPFP=3.3735 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,740B, BPFP=2.4843 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,204B, BPFP=3.3841 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,020B, BPFP=2.4972 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,480B, BPFP=3.3044 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,096B, BPFP=2.1309 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 76,760B, BPFP=3.5484 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 205,324B, BPFP=2.3729 -⌛️ [2/4] FRONTEND: Frontend time: 1.943s (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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.557s - -[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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02650603 4.30296190 - layer.0.v_cache 0.00000028 0.00006265 - layer.1.k_cache 0.00313577 0.35329160 - layer.1.v_cache 0.00000090 0.00020684 - layer.2.k_cache 0.00117709 0.16149258 - layer.2.v_cache 0.00000116 0.00028001 - layer.3.k_cache 0.00129027 0.18141425 - layer.3.v_cache 0.00000225 0.00046378 - layer.4.k_cache 0.00358651 0.45589045 - layer.4.v_cache 0.00000317 0.00074483 - layer.4.output 0.00019489 0.03239859 - ------------------------------------------------------------------------------------- - TOTAL 0.00260593 0.39902880 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 789188 -BPFP 2.6059 bits/point -EBPFP 5.2118 equivalent bits/point -MSE 0.399029 ----------------------- -------------------------------------------------------- -Time: 3.510s Load: 0.010s, Pack+Encode: 1.943s, Decode+Unpack: 1.557s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3990 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample42-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample42-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample43-layer4-item1.zst (56/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample43-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 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, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,532B, BPFP=1.0355 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,016B, BPFP=3.1257 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,860B, BPFP=2.1075 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,532B, BPFP=3.3333 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,284B, BPFP=2.4487 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,480B, BPFP=3.3309 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,404B, BPFP=2.4542 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,108B, BPFP=3.2678 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,384B, BPFP=2.0857 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,536B, BPFP=3.4254 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 191,248B, BPFP=2.1972 -⌛️ [2/4] FRONTEND: Frontend time: 1.962s (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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.532s - -[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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02682906 4.58417251 - layer.0.v_cache 0.00000026 0.00006230 - layer.1.k_cache 0.00324658 0.38314478 - layer.1.v_cache 0.00000081 0.00020213 - layer.2.k_cache 0.00118449 0.15970816 - layer.2.v_cache 0.00000107 0.00027496 - layer.3.k_cache 0.00127207 0.17396824 - layer.3.v_cache 0.00000215 0.00045025 - layer.4.k_cache 0.00355165 0.43918161 - layer.4.v_cache 0.00000309 0.00077282 - layer.4.output 0.00015524 0.02383259 - ------------------------------------------------------------------------------------- - TOTAL 0.00262230 0.41694772 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 770384 -BPFP 2.5288 bits/point -EBPFP 5.0577 equivalent bits/point -MSE 0.416948 ----------------------- -------------------------------------------------------- -Time: 3.504s Load: 0.009s, Pack+Encode: 1.962s, Decode+Unpack: 1.532s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4169 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample43-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample43-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample44-layer4-item1.zst (57/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample44-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 194, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 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, 194, 128) -Output shape: (1, 194, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.0.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.1.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.1.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.2.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.2.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.3.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.3.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.4.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.4.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.4.output: torch.Size([1, 194, 4096]) -> torch.Size([1, 1, 194, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,768B, BPFP=1.0377 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 81,668B, BPFP=3.2888 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 49,848B, BPFP=2.0074 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 87,156B, BPFP=3.5098 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 61,436B, BPFP=2.4741 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 88,976B, BPFP=3.5831 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 62,776B, BPFP=2.5280 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 87,332B, BPFP=3.5169 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 50,260B, BPFP=2.0240 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 90,488B, BPFP=3.6440 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 226,924B, BPFP=2.2846 -⌛️ [2/4] FRONTEND: Frontend time: 2.379s (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, 194, 128]) - layer.0.v_cache: torch.Size([1, 8, 194, 128]) - layer.1.k_cache: torch.Size([1, 8, 194, 128]) - layer.1.v_cache: torch.Size([1, 8, 194, 128]) - layer.2.k_cache: torch.Size([1, 8, 194, 128]) - layer.2.v_cache: torch.Size([1, 8, 194, 128]) - layer.3.k_cache: torch.Size([1, 8, 194, 128]) - layer.3.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.k_cache: torch.Size([1, 8, 194, 128]) - layer.4.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.output: torch.Size([1, 194, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.930s - -[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, 194, 128]) - layer.0.v_cache: torch.Size([1, 8, 194, 128]) - layer.1.k_cache: torch.Size([1, 8, 194, 128]) - layer.1.v_cache: torch.Size([1, 8, 194, 128]) - layer.2.k_cache: torch.Size([1, 8, 194, 128]) - layer.2.v_cache: torch.Size([1, 8, 194, 128]) - layer.3.k_cache: torch.Size([1, 8, 194, 128]) - layer.3.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.k_cache: torch.Size([1, 8, 194, 128]) - layer.4.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.output: torch.Size([1, 194, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02710510 4.43479652 - layer.0.v_cache 0.00000027 0.00006116 - layer.1.k_cache 0.00315558 0.31204711 - layer.1.v_cache 0.00000085 0.00019851 - layer.2.k_cache 0.00115955 0.15824317 - layer.2.v_cache 0.00000114 0.00027008 - layer.3.k_cache 0.00127895 0.17132997 - layer.3.v_cache 0.00000244 0.00045475 - layer.4.k_cache 0.00355671 0.39518824 - layer.4.v_cache 0.00000315 0.00072942 - layer.4.output 0.00014532 0.02575742 - ------------------------------------------------------------------------------------- - TOTAL 0.00263179 0.39831062 - (elements=2,781,184) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2781184 -Total Bytes 912632 -BPFP 2.6252 bits/point -EBPFP 5.2503 equivalent bits/point -MSE 0.398311 ----------------------- -------------------------------------------------------- -Time: 4.323s Load: 0.015s, Pack+Encode: 2.379s, Decode+Unpack: 1.930s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 194, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3983 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample44-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample44-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample45-layer4-item1.zst (58/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample45-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 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, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,716B, BPFP=1.0439 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,120B, BPFP=3.1305 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,724B, BPFP=2.1013 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 77,644B, BPFP=3.5682 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,048B, BPFP=2.4379 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 80,156B, BPFP=3.6836 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,056B, BPFP=2.4842 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,772B, BPFP=3.2983 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,460B, BPFP=2.1351 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,284B, BPFP=3.4138 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 208,788B, BPFP=2.3988 -⌛️ [2/4] FRONTEND: Frontend time: 1.948s (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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.628s - -[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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02613039 4.66537368 - layer.0.v_cache 0.00000027 0.00006954 - layer.1.k_cache 0.00316806 0.42039476 - layer.1.v_cache 0.00000083 0.00021802 - layer.2.k_cache 0.00117426 0.16863227 - layer.2.v_cache 0.00000117 0.00030486 - layer.3.k_cache 0.00130856 0.18193988 - layer.3.v_cache 0.00000223 0.00050865 - layer.4.k_cache 0.00356532 0.49043714 - layer.4.v_cache 0.00000295 0.00070207 - layer.4.output 0.00018112 0.03184550 - ------------------------------------------------------------------------------------- - TOTAL 0.00257704 0.43256878 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 802768 -BPFP 2.6351 bits/point -EBPFP 5.2703 equivalent bits/point -MSE 0.432569 ----------------------- -------------------------------------------------------- -Time: 3.586s Load: 0.010s, Pack+Encode: 1.948s, Decode+Unpack: 1.628s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4326 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample45-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample45-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample46-layer4-item1.zst (59/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample46-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 198, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 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, 198, 128) -Output shape: (1, 198, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.0.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.1.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.1.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.2.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.2.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.3.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.3.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.4.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.4.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.4.output: torch.Size([1, 198, 4096]) -> torch.Size([1, 1, 198, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,292B, BPFP=0.9585 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 83,336B, BPFP=3.2882 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 51,900B, BPFP=2.0478 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 88,888B, BPFP=3.5073 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 62,776B, BPFP=2.4770 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 94,020B, BPFP=3.7098 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 63,928B, BPFP=2.5224 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 88,080B, BPFP=3.4754 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 51,628B, BPFP=2.0371 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 92,024B, BPFP=3.6310 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 220,588B, BPFP=2.1759 -⌛️ [2/4] FRONTEND: Frontend time: 2.374s (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, 198, 128]) - layer.0.v_cache: torch.Size([1, 8, 198, 128]) - layer.1.k_cache: torch.Size([1, 8, 198, 128]) - layer.1.v_cache: torch.Size([1, 8, 198, 128]) - layer.2.k_cache: torch.Size([1, 8, 198, 128]) - layer.2.v_cache: torch.Size([1, 8, 198, 128]) - layer.3.k_cache: torch.Size([1, 8, 198, 128]) - layer.3.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.k_cache: torch.Size([1, 8, 198, 128]) - layer.4.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.output: torch.Size([1, 198, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.008s - -[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, 198, 128]) - layer.0.v_cache: torch.Size([1, 8, 198, 128]) - layer.1.k_cache: torch.Size([1, 8, 198, 128]) - layer.1.v_cache: torch.Size([1, 8, 198, 128]) - layer.2.k_cache: torch.Size([1, 8, 198, 128]) - layer.2.v_cache: torch.Size([1, 8, 198, 128]) - layer.3.k_cache: torch.Size([1, 8, 198, 128]) - layer.3.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.k_cache: torch.Size([1, 8, 198, 128]) - layer.4.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.output: torch.Size([1, 198, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02819830 4.12428854 - layer.0.v_cache 0.00000029 0.00006093 - layer.1.k_cache 0.00303339 0.32512599 - layer.1.v_cache 0.00000075 0.00018093 - layer.2.k_cache 0.00117903 0.16032812 - layer.2.v_cache 0.00000119 0.00025942 - layer.3.k_cache 0.00132068 0.17765903 - layer.3.v_cache 0.00000222 0.00044293 - layer.4.k_cache 0.00360013 0.42255856 - layer.4.v_cache 0.00000329 0.00070217 - layer.4.output 0.00021245 0.03357726 - ------------------------------------------------------------------------------------- - TOTAL 0.00272779 0.38185112 - (elements=2,838,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2838528 -Total Bytes 921460 -BPFP 2.5970 bits/point -EBPFP 5.1940 equivalent bits/point -MSE 0.381851 ----------------------- -------------------------------------------------------- -Time: 4.393s Load: 0.011s, Pack+Encode: 2.374s, Decode+Unpack: 2.008s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 198, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3819 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample46-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample46-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample48-layer4-item1.zst (60/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample48-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: 0.009s - ------------------------------------------------------------- -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: 24,428B, BPFP=1.1031 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,704B, BPFP=3.0574 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,792B, BPFP=2.1131 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,676B, BPFP=3.2820 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,548B, BPFP=2.4633 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,160B, BPFP=3.3941 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,500B, BPFP=2.4612 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,612B, BPFP=3.2339 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,016B, BPFP=2.1232 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 75,900B, BPFP=3.4276 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 205,360B, BPFP=2.3185 -⌛️ [2/4] FRONTEND: Frontend time: 1.975s (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: 1.564s - -[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 0.02631039 4.81601146 - layer.0.v_cache 0.00000026 0.00005954 - layer.1.k_cache 0.00315605 0.43069978 - layer.1.v_cache 0.00000083 0.00020096 - layer.2.k_cache 0.00121195 0.16685220 - layer.2.v_cache 0.00000119 0.00029642 - layer.3.k_cache 0.00131129 0.17651233 - layer.3.v_cache 0.00000228 0.00047413 - layer.4.k_cache 0.00352345 0.43602268 - layer.4.v_cache 0.00000333 0.00079627 - layer.4.output 0.00017266 0.03111714 - ------------------------------------------------------------------------------------- - TOTAL 0.00258655 0.43945674 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 795696 -BPFP 2.5666 bits/point -EBPFP 5.1333 equivalent bits/point -MSE 0.439457 ----------------------- -------------------------------------------------------- -Time: 3.548s Load: 0.009s, Pack+Encode: 1.975s, Decode+Unpack: 1.564s ----------------------- -------------------------------------------------------- -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 0.4395 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample48-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample48-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample49-layer4-item1.zst (61/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample49-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 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, 175, 128) -Output shape: (1, 175, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.output: torch.Size([1, 175, 4096]) -> torch.Size([1, 1, 175, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,348B, BPFP=1.0870 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,996B, BPFP=3.0802 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,504B, BPFP=2.0761 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 73,340B, BPFP=3.2741 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,336B, BPFP=2.4257 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,160B, BPFP=3.2661 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,500B, BPFP=2.4330 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,472B, BPFP=3.1907 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,704B, BPFP=2.0850 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,500B, BPFP=3.3259 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 207,416B, BPFP=2.3149 -⌛️ [2/4] FRONTEND: Frontend time: 1.967s (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, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.563s - -[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, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03100830 4.30551967 - layer.0.v_cache 0.00000027 0.00006013 - layer.1.k_cache 0.00306070 0.40167306 - layer.1.v_cache 0.00000080 0.00019247 - layer.2.k_cache 0.00119082 0.16403309 - layer.2.v_cache 0.00000118 0.00028069 - layer.3.k_cache 0.00134761 0.18195338 - layer.3.v_cache 0.00000229 0.00045783 - layer.4.k_cache 0.00348653 0.45714643 - layer.4.v_cache 0.00000313 0.00069188 - layer.4.output 0.00023428 0.03834999 - ------------------------------------------------------------------------------------- - TOTAL 0.00293134 0.40467204 - (elements=2,508,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2508800 -Total Bytes 795276 -BPFP 2.5360 bits/point -EBPFP 5.0719 equivalent bits/point -MSE 0.404672 ----------------------- -------------------------------------------------------- -Time: 3.540s Load: 0.010s, Pack+Encode: 1.967s, Decode+Unpack: 1.563s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4047 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample49-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample49-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample5-layer4-item1.zst (62/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample5-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 240, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.014s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 240, 128) -Output shape: (1, 240, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.0.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.1.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.1.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.2.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.2.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.3.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.3.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.4.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.4.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.4.output: torch.Size([1, 240, 4096]) -> torch.Size([1, 1, 240, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 34,760B, BPFP=1.1315 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 91,448B, BPFP=2.9768 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 60,944B, BPFP=1.9839 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 99,860B, BPFP=3.2507 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 72,012B, BPFP=2.3441 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 101,988B, BPFP=3.3199 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 72,792B, BPFP=2.3695 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 95,272B, BPFP=3.1013 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 61,892B, BPFP=2.0147 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 100,228B, BPFP=3.2626 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 290,768B, BPFP=2.3663 -⌛️ [2/4] FRONTEND: Frontend time: 2.379s (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, 240, 128]) - layer.0.v_cache: torch.Size([1, 8, 240, 128]) - layer.1.k_cache: torch.Size([1, 8, 240, 128]) - layer.1.v_cache: torch.Size([1, 8, 240, 128]) - layer.2.k_cache: torch.Size([1, 8, 240, 128]) - layer.2.v_cache: torch.Size([1, 8, 240, 128]) - layer.3.k_cache: torch.Size([1, 8, 240, 128]) - layer.3.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.k_cache: torch.Size([1, 8, 240, 128]) - layer.4.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.output: torch.Size([1, 240, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.025s - -[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, 240, 128]) - layer.0.v_cache: torch.Size([1, 8, 240, 128]) - layer.1.k_cache: torch.Size([1, 8, 240, 128]) - layer.1.v_cache: torch.Size([1, 8, 240, 128]) - layer.2.k_cache: torch.Size([1, 8, 240, 128]) - layer.2.v_cache: torch.Size([1, 8, 240, 128]) - layer.3.k_cache: torch.Size([1, 8, 240, 128]) - layer.3.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.k_cache: torch.Size([1, 8, 240, 128]) - layer.4.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.output: torch.Size([1, 240, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02660088 5.80351613 - layer.0.v_cache 0.00000028 0.00005916 - layer.1.k_cache 0.00295962 0.37473068 - layer.1.v_cache 0.00000087 0.00019722 - layer.2.k_cache 0.00121828 0.15918689 - layer.2.v_cache 0.00000120 0.00026732 - layer.3.k_cache 0.00127180 0.16632578 - layer.3.v_cache 0.00000260 0.00045106 - layer.4.k_cache 0.00365680 0.42208541 - layer.4.v_cache 0.00000312 0.00070443 - layer.4.output 0.00015244 0.02752598 - ------------------------------------------------------------------------------------- - TOTAL 0.00259466 0.50268771 - (elements=3,440,640) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3440640 -Total Bytes 1081964 -BPFP 2.5157 bits/point -EBPFP 5.0315 equivalent bits/point -MSE 0.502688 ----------------------- -------------------------------------------------------- -Time: 4.417s Load: 0.014s, Pack+Encode: 2.379s, Decode+Unpack: 2.025s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 240, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.5027 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample5-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample5-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample50-layer4-item1.zst (63/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample50-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 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, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,444B, BPFP=1.0518 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 71,588B, BPFP=2.9592 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,800B, BPFP=1.9345 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 78,508B, BPFP=3.2452 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 55,380B, BPFP=2.2892 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 80,500B, BPFP=3.3275 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,812B, BPFP=2.2657 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 73,528B, BPFP=3.0394 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,340B, BPFP=1.9568 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 80,380B, BPFP=3.3226 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 223,380B, BPFP=2.3084 -⌛️ [2/4] FRONTEND: Frontend time: 1.976s (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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.561s - -[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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02646759 4.35515227 - layer.0.v_cache 0.00000027 0.00006312 - layer.1.k_cache 0.00304090 0.42307625 - layer.1.v_cache 0.00000085 0.00020625 - layer.2.k_cache 0.00115999 0.15492610 - layer.2.v_cache 0.00000120 0.00029004 - layer.3.k_cache 0.00125869 0.16816461 - layer.3.v_cache 0.00000235 0.00046613 - layer.4.k_cache 0.00349331 0.41783332 - layer.4.v_cache 0.00000329 0.00074305 - layer.4.output 0.00016341 0.03361693 - ------------------------------------------------------------------------------------- - TOTAL 0.00257729 0.40395635 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 837660 -BPFP 2.4732 bits/point -EBPFP 4.9465 equivalent bits/point -MSE 0.403956 ----------------------- -------------------------------------------------------- -Time: 3.547s Load: 0.010s, Pack+Encode: 1.976s, Decode+Unpack: 1.561s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4040 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample50-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample50-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample51-layer4-item1.zst (64/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample51-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 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, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,004B, BPFP=1.1031 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,588B, BPFP=3.1520 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,396B, BPFP=2.0862 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,508B, BPFP=3.2862 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,656B, BPFP=2.4658 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,608B, BPFP=3.4287 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,880B, BPFP=2.4761 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,044B, BPFP=3.2649 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,208B, BPFP=2.1235 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,964B, BPFP=3.3531 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 202,404B, BPFP=2.3254 -⌛️ [2/4] FRONTEND: Frontend time: 1.973s (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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.575s - -[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, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03233885 5.06633229 - layer.0.v_cache 0.00000026 0.00006011 - layer.1.k_cache 0.00307676 0.34438542 - layer.1.v_cache 0.00000080 0.00019696 - layer.2.k_cache 0.00113683 0.16080247 - layer.2.v_cache 0.00000124 0.00028322 - layer.3.k_cache 0.00130099 0.17555669 - layer.3.v_cache 0.00000223 0.00046234 - layer.4.k_cache 0.00352303 0.43075530 - layer.4.v_cache 0.00000334 0.00074115 - layer.4.output 0.00016991 0.03397415 - ------------------------------------------------------------------------------------- - TOTAL 0.00300457 0.45110518 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 784260 -BPFP 2.5744 bits/point -EBPFP 5.1488 equivalent bits/point -MSE 0.451105 ----------------------- -------------------------------------------------------- -Time: 3.557s Load: 0.010s, Pack+Encode: 1.973s, Decode+Unpack: 1.575s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4511 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample51-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample51-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample52-layer4-item1.zst (65/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample52-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 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, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,564B, BPFP=1.0663 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,804B, BPFP=3.3032 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,984B, BPFP=2.0759 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 69,972B, BPFP=3.4598 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 51,116B, BPFP=2.5275 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,136B, BPFP=3.5174 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,960B, BPFP=2.5198 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 69,720B, BPFP=3.4474 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,444B, BPFP=2.0987 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,696B, BPFP=3.5451 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 190,364B, BPFP=2.3532 -⌛️ [2/4] FRONTEND: Frontend time: 1.971s (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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.531s - -[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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02846188 4.37558640 - layer.0.v_cache 0.00000028 0.00006245 - layer.1.k_cache 0.00321952 0.32988843 - layer.1.v_cache 0.00000089 0.00020493 - layer.2.k_cache 0.00116433 0.16851871 - layer.2.v_cache 0.00000127 0.00028471 - layer.3.k_cache 0.00135431 0.18366550 - layer.3.v_cache 0.00000232 0.00046548 - layer.4.k_cache 0.00338441 0.43607306 - layer.4.v_cache 0.00000314 0.00072864 - layer.4.output 0.00020823 0.03949726 - ------------------------------------------------------------------------------------- - TOTAL 0.00274466 0.40381910 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 747760 -BPFP 2.6410 bits/point -EBPFP 5.2820 equivalent bits/point -MSE 0.403819 ----------------------- -------------------------------------------------------- -Time: 3.511s Load: 0.009s, Pack+Encode: 1.971s, Decode+Unpack: 1.531s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4038 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample52-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample52-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample53-layer4-item1.zst (66/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample53-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 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, 185, 128) -Output shape: (1, 185, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.output: torch.Size([1, 185, 4096]) -> torch.Size([1, 1, 185, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,688B, BPFP=1.0848 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,412B, BPFP=2.8890 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,904B, BPFP=1.9807 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,484B, BPFP=3.1454 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,956B, BPFP=2.3208 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,048B, BPFP=3.1693 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,760B, BPFP=2.3125 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,736B, BPFP=3.0294 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,592B, BPFP=2.0098 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,300B, BPFP=3.1377 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 208,400B, BPFP=2.2002 -⌛️ [2/4] FRONTEND: Frontend time: 1.982s (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, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.533s - -[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, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02684415 4.61155049 - layer.0.v_cache 0.00000026 0.00006108 - layer.1.k_cache 0.00308337 0.42129261 - layer.1.v_cache 0.00000086 0.00019546 - layer.2.k_cache 0.00119263 0.16006123 - layer.2.v_cache 0.00000115 0.00027595 - layer.3.k_cache 0.00133167 0.18099953 - layer.3.v_cache 0.00000224 0.00047257 - layer.4.k_cache 0.00346133 0.43700050 - layer.4.v_cache 0.00000321 0.00077282 - layer.4.output 0.00019684 0.03026630 - ------------------------------------------------------------------------------------- - TOTAL 0.00262202 0.42383910 - (elements=2,652,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2652160 -Total Bytes 802280 -BPFP 2.4200 bits/point -EBPFP 4.8400 equivalent bits/point -MSE 0.423839 ----------------------- -------------------------------------------------------- -Time: 3.526s Load: 0.011s, Pack+Encode: 1.982s, Decode+Unpack: 1.533s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4238 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample53-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample53-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample54-layer4-item1.zst (67/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample54-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 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, 165, 128) -Output shape: (1, 165, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.output: torch.Size([1, 165, 4096]) -> torch.Size([1, 1, 165, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,704B, BPFP=1.0750 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,984B, BPFP=3.2189 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,144B, BPFP=2.0902 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,792B, BPFP=3.3992 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 52,964B, BPFP=2.5078 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,936B, BPFP=3.4534 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,304B, BPFP=2.5239 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,260B, BPFP=3.3741 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,484B, BPFP=2.1536 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,836B, BPFP=3.4487 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 198,764B, BPFP=2.3528 -⌛️ [2/4] FRONTEND: Frontend time: 1.976s (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, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.553s - -[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, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02722176 4.11112985 - layer.0.v_cache 0.00000026 0.00005469 - layer.1.k_cache 0.00312822 0.33676041 - layer.1.v_cache 0.00000086 0.00019601 - layer.2.k_cache 0.00119859 0.16510144 - layer.2.v_cache 0.00000118 0.00026638 - layer.3.k_cache 0.00130010 0.17408046 - layer.3.v_cache 0.00000225 0.00043528 - layer.4.k_cache 0.00349983 0.43692798 - layer.4.v_cache 0.00000307 0.00070408 - layer.4.output 0.00018894 0.03539944 - ------------------------------------------------------------------------------------- - TOTAL 0.00265085 0.38337531 - (elements=2,365,440) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2365440 -Total Bytes 774172 -BPFP 2.6183 bits/point -EBPFP 5.2366 equivalent bits/point -MSE 0.383375 ----------------------- -------------------------------------------------------- -Time: 3.538s Load: 0.009s, Pack+Encode: 1.976s, Decode+Unpack: 1.553s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3834 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample54-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample54-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample56-layer4-item1.zst (68/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample56-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 210, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 210, 128) -Output shape: (1, 210, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.0.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.1.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.1.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.2.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.2.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.3.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.3.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.4.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.4.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.4.output: torch.Size([1, 210, 4096]) -> torch.Size([1, 1, 210, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,832B, BPFP=0.9610 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 85,892B, BPFP=3.1954 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 54,164B, BPFP=2.0150 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 91,352B, BPFP=3.3985 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 65,468B, BPFP=2.4356 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 92,936B, BPFP=3.4574 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 66,916B, BPFP=2.4894 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 90,588B, BPFP=3.3701 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 54,088B, BPFP=2.0122 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 93,924B, BPFP=3.4942 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 232,904B, BPFP=2.1661 -⌛️ [2/4] FRONTEND: Frontend time: 2.372s (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, 210, 128]) - layer.0.v_cache: torch.Size([1, 8, 210, 128]) - layer.1.k_cache: torch.Size([1, 8, 210, 128]) - layer.1.v_cache: torch.Size([1, 8, 210, 128]) - layer.2.k_cache: torch.Size([1, 8, 210, 128]) - layer.2.v_cache: torch.Size([1, 8, 210, 128]) - layer.3.k_cache: torch.Size([1, 8, 210, 128]) - layer.3.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.k_cache: torch.Size([1, 8, 210, 128]) - layer.4.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.output: torch.Size([1, 210, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.035s - -[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, 210, 128]) - layer.0.v_cache: torch.Size([1, 8, 210, 128]) - layer.1.k_cache: torch.Size([1, 8, 210, 128]) - layer.1.v_cache: torch.Size([1, 8, 210, 128]) - layer.2.k_cache: torch.Size([1, 8, 210, 128]) - layer.2.v_cache: torch.Size([1, 8, 210, 128]) - layer.3.k_cache: torch.Size([1, 8, 210, 128]) - layer.3.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.k_cache: torch.Size([1, 8, 210, 128]) - layer.4.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.output: torch.Size([1, 210, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02702481 4.45866263 - layer.0.v_cache 0.00000029 0.00005945 - layer.1.k_cache 0.00301925 0.32568378 - layer.1.v_cache 0.00000076 0.00018058 - layer.2.k_cache 0.00118472 0.16792003 - layer.2.v_cache 0.00000128 0.00024709 - layer.3.k_cache 0.00128382 0.17252063 - layer.3.v_cache 0.00000210 0.00041640 - layer.4.k_cache 0.00361221 0.45032385 - layer.4.v_cache 0.00000301 0.00066219 - layer.4.output 0.00015169 0.02998874 - ------------------------------------------------------------------------------------- - TOTAL 0.00262422 0.40690226 - (elements=3,010,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3010560 -Total Bytes 954064 -BPFP 2.5352 bits/point -EBPFP 5.0705 equivalent bits/point -MSE 0.406902 ----------------------- -------------------------------------------------------- -Time: 4.419s Load: 0.012s, Pack+Encode: 2.372s, Decode+Unpack: 2.035s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 210, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4069 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample56-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample56-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample57-layer4-item1.zst (69/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample57-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 146, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 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, 146, 128) -Output shape: (1, 146, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.0.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.1.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.1.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.2.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.2.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.3.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.3.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.4.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.4.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.4.output: torch.Size([1, 146, 4096]) -> torch.Size([1, 1, 146, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 20,432B, BPFP=1.0933 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 64,164B, BPFP=3.4334 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 39,948B, BPFP=2.1376 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 68,364B, BPFP=3.6582 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 49,548B, BPFP=2.6513 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 69,296B, BPFP=3.7080 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 49,508B, BPFP=2.6492 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,036B, BPFP=3.6406 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 40,952B, BPFP=2.1914 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 69,436B, BPFP=3.7155 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 193,520B, BPFP=2.5888 -⌛️ [2/4] FRONTEND: Frontend time: 1.972s (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, 146, 128]) - layer.0.v_cache: torch.Size([1, 8, 146, 128]) - layer.1.k_cache: torch.Size([1, 8, 146, 128]) - layer.1.v_cache: torch.Size([1, 8, 146, 128]) - layer.2.k_cache: torch.Size([1, 8, 146, 128]) - layer.2.v_cache: torch.Size([1, 8, 146, 128]) - layer.3.k_cache: torch.Size([1, 8, 146, 128]) - layer.3.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.k_cache: torch.Size([1, 8, 146, 128]) - layer.4.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.output: torch.Size([1, 146, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.526s - -[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, 146, 128]) - layer.0.v_cache: torch.Size([1, 8, 146, 128]) - layer.1.k_cache: torch.Size([1, 8, 146, 128]) - layer.1.v_cache: torch.Size([1, 8, 146, 128]) - layer.2.k_cache: torch.Size([1, 8, 146, 128]) - layer.2.v_cache: torch.Size([1, 8, 146, 128]) - layer.3.k_cache: torch.Size([1, 8, 146, 128]) - layer.3.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.k_cache: torch.Size([1, 8, 146, 128]) - layer.4.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.output: torch.Size([1, 146, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02740652 4.47998799 - layer.0.v_cache 0.00000027 0.00006334 - layer.1.k_cache 0.00310827 0.33562169 - layer.1.v_cache 0.00000089 0.00021511 - layer.2.k_cache 0.00118257 0.16228030 - layer.2.v_cache 0.00000116 0.00029351 - layer.3.k_cache 0.00130374 0.18173125 - layer.3.v_cache 0.00000225 0.00048232 - layer.4.k_cache 0.00337717 0.39173226 - layer.4.v_cache 0.00000316 0.00079843 - layer.4.output 0.00015578 0.02850596 - ------------------------------------------------------------------------------------- - TOTAL 0.00264351 0.40480214 - (elements=2,093,056) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2093056 -Total Bytes 733204 -BPFP 2.8024 bits/point -EBPFP 5.6048 equivalent bits/point -MSE 0.404802 ----------------------- -------------------------------------------------------- -Time: 3.506s Load: 0.008s, Pack+Encode: 1.972s, Decode+Unpack: 1.526s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 146, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4048 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample57-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample57-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample6-layer4-item1.zst (70/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample6-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 223, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 223, 128) -Output shape: (1, 223, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.0.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.1.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.1.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.2.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.2.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.3.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.3.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.4.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.4.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.4.output: torch.Size([1, 223, 4096]) -> torch.Size([1, 1, 223, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 30,168B, BPFP=1.0569 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 90,048B, BPFP=3.1547 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 57,108B, BPFP=2.0007 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 95,764B, BPFP=3.3550 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 69,676B, BPFP=2.4410 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 96,032B, BPFP=3.3643 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 69,720B, BPFP=2.4425 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 94,312B, BPFP=3.3041 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 57,828B, BPFP=2.0259 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 97,152B, BPFP=3.4036 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 264,460B, BPFP=2.3162 -⌛️ [2/4] FRONTEND: Frontend time: 2.400s (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, 223, 128]) - layer.0.v_cache: torch.Size([1, 8, 223, 128]) - layer.1.k_cache: torch.Size([1, 8, 223, 128]) - layer.1.v_cache: torch.Size([1, 8, 223, 128]) - layer.2.k_cache: torch.Size([1, 8, 223, 128]) - layer.2.v_cache: torch.Size([1, 8, 223, 128]) - layer.3.k_cache: torch.Size([1, 8, 223, 128]) - layer.3.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.k_cache: torch.Size([1, 8, 223, 128]) - layer.4.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.output: torch.Size([1, 223, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.072s - -[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, 223, 128]) - layer.0.v_cache: torch.Size([1, 8, 223, 128]) - layer.1.k_cache: torch.Size([1, 8, 223, 128]) - layer.1.v_cache: torch.Size([1, 8, 223, 128]) - layer.2.k_cache: torch.Size([1, 8, 223, 128]) - layer.2.v_cache: torch.Size([1, 8, 223, 128]) - layer.3.k_cache: torch.Size([1, 8, 223, 128]) - layer.3.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.k_cache: torch.Size([1, 8, 223, 128]) - layer.4.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.output: torch.Size([1, 223, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02710374 4.46663019 - layer.0.v_cache 0.00000027 0.00005906 - layer.1.k_cache 0.00293174 0.31988577 - layer.1.v_cache 0.00000088 0.00020129 - layer.2.k_cache 0.00135432 0.16126121 - layer.2.v_cache 0.00000122 0.00027193 - layer.3.k_cache 0.00128520 0.16607733 - layer.3.v_cache 0.00000265 0.00047992 - layer.4.k_cache 0.00356170 0.39409008 - layer.4.v_cache 0.00000337 0.00070916 - layer.4.output 0.00016975 0.03105625 - ------------------------------------------------------------------------------------- - TOTAL 0.00263743 0.40242078 - (elements=3,196,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3196928 -Total Bytes 1022268 -BPFP 2.5581 bits/point -EBPFP 5.1163 equivalent bits/point -MSE 0.402421 ----------------------- -------------------------------------------------------- -Time: 4.484s Load: 0.012s, Pack+Encode: 2.400s, Decode+Unpack: 2.072s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 223, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4024 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample6-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample6-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample60-layer4-item1.zst (71/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample60-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 167, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 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, 167, 128) -Output shape: (1, 167, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.0.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.1.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.1.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.2.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.2.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.3.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.3.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.4.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.4.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.4.output: torch.Size([1, 167, 4096]) -> torch.Size([1, 1, 167, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,668B, BPFP=1.0604 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,440B, BPFP=3.2017 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,984B, BPFP=2.1044 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,288B, BPFP=3.3817 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,280B, BPFP=2.4925 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,676B, BPFP=3.3999 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,408B, BPFP=2.4985 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,116B, BPFP=3.3269 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,692B, BPFP=2.1375 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,664B, BPFP=3.3993 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 199,316B, BPFP=2.3311 -⌛️ [2/4] FRONTEND: Frontend time: 1.947s (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, 167, 128]) - layer.0.v_cache: torch.Size([1, 8, 167, 128]) - layer.1.k_cache: torch.Size([1, 8, 167, 128]) - layer.1.v_cache: torch.Size([1, 8, 167, 128]) - layer.2.k_cache: torch.Size([1, 8, 167, 128]) - layer.2.v_cache: torch.Size([1, 8, 167, 128]) - layer.3.k_cache: torch.Size([1, 8, 167, 128]) - layer.3.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.k_cache: torch.Size([1, 8, 167, 128]) - layer.4.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.output: torch.Size([1, 167, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.581s - -[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, 167, 128]) - layer.0.v_cache: torch.Size([1, 8, 167, 128]) - layer.1.k_cache: torch.Size([1, 8, 167, 128]) - layer.1.v_cache: torch.Size([1, 8, 167, 128]) - layer.2.k_cache: torch.Size([1, 8, 167, 128]) - layer.2.v_cache: torch.Size([1, 8, 167, 128]) - layer.3.k_cache: torch.Size([1, 8, 167, 128]) - layer.3.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.k_cache: torch.Size([1, 8, 167, 128]) - layer.4.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.output: torch.Size([1, 167, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02821272 4.18786329 - layer.0.v_cache 0.00000027 0.00005877 - layer.1.k_cache 0.00309555 0.34649069 - layer.1.v_cache 0.00000088 0.00020272 - layer.2.k_cache 0.00116749 0.16042714 - layer.2.v_cache 0.00000113 0.00027129 - layer.3.k_cache 0.00133544 0.18890774 - layer.3.v_cache 0.00000222 0.00045911 - layer.4.k_cache 0.00347765 0.45854877 - layer.4.v_cache 0.00000297 0.00070621 - layer.4.output 0.00020541 0.03381020 - ------------------------------------------------------------------------------------- - TOTAL 0.00272271 0.39136975 - (elements=2,394,112) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2394112 -Total Bytes 776532 -BPFP 2.5948 bits/point -EBPFP 5.1896 equivalent bits/point -MSE 0.391370 ----------------------- -------------------------------------------------------- -Time: 3.537s Load: 0.010s, Pack+Encode: 1.947s, Decode+Unpack: 1.581s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 167, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3914 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample60-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample60-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample61-layer4-item1.zst (72/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample61-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 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, 162, 128) -Output shape: (1, 162, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.output: torch.Size([1, 162, 4096]) -> torch.Size([1, 1, 162, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,488B, BPFP=1.0845 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,728B, BPFP=3.3144 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 43,672B, BPFP=2.1061 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,640B, BPFP=3.4549 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 52,020B, BPFP=2.5087 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,600B, BPFP=3.5012 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 52,552B, BPFP=2.5343 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,288B, BPFP=3.4379 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 44,616B, BPFP=2.1516 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,064B, BPFP=3.5235 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 200,472B, BPFP=2.4170 -⌛️ [2/4] FRONTEND: Frontend time: 1.962s (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, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.539s - -[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, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02804119 4.51550557 - layer.0.v_cache 0.00000028 0.00006157 - layer.1.k_cache 0.00315224 0.36125442 - layer.1.v_cache 0.00000083 0.00020832 - layer.2.k_cache 0.00116199 0.16258739 - layer.2.v_cache 0.00000117 0.00027682 - layer.3.k_cache 0.00129529 0.17944483 - layer.3.v_cache 0.00000233 0.00048386 - layer.4.k_cache 0.00343653 0.43083356 - layer.4.v_cache 0.00000316 0.00075884 - layer.4.output 0.00017336 0.02769988 - ------------------------------------------------------------------------------------- - TOTAL 0.00269917 0.41158677 - (elements=2,322,432) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2322432 -Total Bytes 773140 -BPFP 2.6632 bits/point -EBPFP 5.3264 equivalent bits/point -MSE 0.411587 ----------------------- -------------------------------------------------------- -Time: 3.511s Load: 0.010s, Pack+Encode: 1.962s, Decode+Unpack: 1.539s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4116 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample61-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample61-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample62-layer4-item1.zst (73/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample62-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 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, 176, 128) -Output shape: (1, 176, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.output: torch.Size([1, 176, 4096]) -> torch.Size([1, 1, 176, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,356B, BPFP=1.1255 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,500B, BPFP=3.0407 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,584B, BPFP=2.0678 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,780B, BPFP=3.2306 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,548B, BPFP=2.4213 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 76,096B, BPFP=3.3778 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,540B, BPFP=2.4210 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,600B, BPFP=3.2227 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,788B, BPFP=2.0769 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,736B, BPFP=3.3175 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 222,328B, BPFP=2.4672 -⌛️ [2/4] FRONTEND: Frontend time: 1.955s (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, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.540s - -[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, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02849088 4.53333664 - layer.0.v_cache 0.00000028 0.00006035 - layer.1.k_cache 0.00305327 0.40221197 - layer.1.v_cache 0.00000091 0.00019957 - layer.2.k_cache 0.00120035 0.16046029 - layer.2.v_cache 0.00000116 0.00026976 - layer.3.k_cache 0.00128172 0.17278965 - layer.3.v_cache 0.00000269 0.00046083 - layer.4.k_cache 0.00351652 0.42848804 - layer.4.v_cache 0.00000314 0.00071183 - layer.4.output 0.00018003 0.03118321 - ------------------------------------------------------------------------------------- - TOTAL 0.00273365 0.41598013 - (elements=2,523,136) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2523136 -Total Bytes 814856 -BPFP 2.5836 bits/point -EBPFP 5.1673 equivalent bits/point -MSE 0.415980 ----------------------- -------------------------------------------------------- -Time: 3.505s Load: 0.011s, Pack+Encode: 1.955s, Decode+Unpack: 1.540s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4160 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample62-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample62-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample64-layer4-item1.zst (74/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample64-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 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, 168, 128) -Output shape: (1, 168, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.output: torch.Size([1, 168, 4096]) -> torch.Size([1, 1, 168, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,360B, BPFP=1.0863 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,904B, BPFP=3.2042 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,092B, BPFP=2.0969 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,656B, BPFP=3.3787 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,200B, BPFP=2.4740 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,376B, BPFP=3.3657 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,876B, BPFP=2.5054 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,420B, BPFP=3.3212 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,236B, BPFP=2.1501 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,524B, BPFP=3.4191 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 194,140B, BPFP=2.2570 -⌛️ [2/4] FRONTEND: Frontend time: 2.026s (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, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.526s - -[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, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02847742 4.23681495 - layer.0.v_cache 0.00000027 0.00006108 - layer.1.k_cache 0.00308642 0.34853081 - layer.1.v_cache 0.00000083 0.00019572 - layer.2.k_cache 0.00117363 0.16163115 - layer.2.v_cache 0.00000114 0.00025824 - layer.3.k_cache 0.00131494 0.18013695 - layer.3.v_cache 0.00000212 0.00042658 - layer.4.k_cache 0.00375810 0.44582094 - layer.4.v_cache 0.00000313 0.00067944 - layer.4.output 0.00017449 0.03078489 - ------------------------------------------------------------------------------------- - TOTAL 0.00275114 0.39269253 - (elements=2,408,448) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2408448 -Total Bytes 774784 -BPFP 2.5736 bits/point -EBPFP 5.1471 equivalent bits/point -MSE 0.392693 ----------------------- -------------------------------------------------------- -Time: 3.560s Load: 0.009s, Pack+Encode: 2.026s, Decode+Unpack: 1.526s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3927 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample64-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample64-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample66-layer4-item1.zst (75/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample66-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 160, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 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, 160, 128) -Output shape: (1, 160, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.0.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.1.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.1.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.2.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.2.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.3.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.3.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.4.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.4.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.4.output: torch.Size([1, 160, 4096]) -> torch.Size([1, 1, 160, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,220B, BPFP=1.0361 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,340B, BPFP=3.2393 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 42,488B, BPFP=2.0746 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,496B, BPFP=3.4422 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,888B, BPFP=2.4848 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,460B, BPFP=3.4893 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 51,568B, BPFP=2.5180 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,004B, BPFP=3.4182 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,916B, BPFP=2.0955 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,824B, BPFP=3.5070 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 183,320B, BPFP=2.2378 -⌛️ [2/4] FRONTEND: Frontend time: 1.950s (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, 160, 128]) - layer.0.v_cache: torch.Size([1, 8, 160, 128]) - layer.1.k_cache: torch.Size([1, 8, 160, 128]) - layer.1.v_cache: torch.Size([1, 8, 160, 128]) - layer.2.k_cache: torch.Size([1, 8, 160, 128]) - layer.2.v_cache: torch.Size([1, 8, 160, 128]) - layer.3.k_cache: torch.Size([1, 8, 160, 128]) - layer.3.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.k_cache: torch.Size([1, 8, 160, 128]) - layer.4.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.output: torch.Size([1, 160, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.571s - -[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, 160, 128]) - layer.0.v_cache: torch.Size([1, 8, 160, 128]) - layer.1.k_cache: torch.Size([1, 8, 160, 128]) - layer.1.v_cache: torch.Size([1, 8, 160, 128]) - layer.2.k_cache: torch.Size([1, 8, 160, 128]) - layer.2.v_cache: torch.Size([1, 8, 160, 128]) - layer.3.k_cache: torch.Size([1, 8, 160, 128]) - layer.3.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.k_cache: torch.Size([1, 8, 160, 128]) - layer.4.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.output: torch.Size([1, 160, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02762303 4.56138611 - layer.0.v_cache 0.00000027 0.00006188 - layer.1.k_cache 0.00317047 0.33162239 - layer.1.v_cache 0.00000088 0.00019955 - layer.2.k_cache 0.00116413 0.16862855 - layer.2.v_cache 0.00000117 0.00028333 - layer.3.k_cache 0.00133371 0.18735366 - layer.3.v_cache 0.00000211 0.00043511 - layer.4.k_cache 0.00351672 0.44742794 - layer.4.v_cache 0.00000299 0.00070100 - layer.4.output 0.00019762 0.04444838 - ------------------------------------------------------------------------------------- - TOTAL 0.00268614 0.41970665 - (elements=2,293,760) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2293760 -Total Bytes 742524 -BPFP 2.5897 bits/point -EBPFP 5.1794 equivalent bits/point -MSE 0.419707 ----------------------- -------------------------------------------------------- -Time: 3.531s Load: 0.010s, Pack+Encode: 1.950s, Decode+Unpack: 1.571s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 160, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4197 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample66-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample66-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample67-layer4-item1.zst (76/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample67-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 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, 168, 128) -Output shape: (1, 168, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.output: torch.Size([1, 168, 4096]) -> torch.Size([1, 1, 168, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,688B, BPFP=1.0551 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,556B, BPFP=3.1881 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 44,376B, BPFP=2.0636 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,588B, BPFP=3.3756 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,816B, BPFP=2.5026 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,172B, BPFP=3.4027 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,720B, BPFP=2.4981 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,868B, BPFP=3.3421 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 45,760B, BPFP=2.1280 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,592B, BPFP=3.4688 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 198,088B, BPFP=2.3029 -⌛️ [2/4] FRONTEND: Frontend time: 1.957s (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, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.538s - -[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, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03000965 4.87030247 - layer.0.v_cache 0.00000027 0.00006354 - layer.1.k_cache 0.00306367 0.33231554 - layer.1.v_cache 0.00000085 0.00019633 - layer.2.k_cache 0.00121342 0.16135515 - layer.2.v_cache 0.00000122 0.00027561 - layer.3.k_cache 0.00130310 0.17844045 - layer.3.v_cache 0.00000229 0.00046391 - layer.4.k_cache 0.00355794 0.42784278 - layer.4.v_cache 0.00000315 0.00073504 - layer.4.output 0.00017029 0.02927752 - ------------------------------------------------------------------------------------- - TOTAL 0.00284548 0.43493578 - (elements=2,408,448) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2408448 -Total Bytes 779224 -BPFP 2.5883 bits/point -EBPFP 5.1766 equivalent bits/point -MSE 0.434936 ----------------------- -------------------------------------------------------- -Time: 3.504s Load: 0.010s, Pack+Encode: 1.957s, Decode+Unpack: 1.538s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4349 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample67-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample67-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample68-layer4-item1.zst (77/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample68-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 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, 176, 128) -Output shape: (1, 176, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.output: torch.Size([1, 176, 4096]) -> torch.Size([1, 1, 176, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,540B, BPFP=1.1337 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,984B, BPFP=3.0178 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,568B, BPFP=2.0671 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,444B, BPFP=3.1713 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 55,788B, BPFP=2.4764 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,360B, BPFP=3.2564 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,228B, BPFP=2.4071 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,160B, BPFP=3.1587 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,824B, BPFP=2.0785 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,700B, BPFP=3.2715 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 201,384B, BPFP=2.2348 -⌛️ [2/4] FRONTEND: Frontend time: 1.981s (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, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.563s - -[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, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02702593 5.12016817 - layer.0.v_cache 0.00000026 0.00005821 - layer.1.k_cache 0.00320345 0.42895135 - layer.1.v_cache 0.00000081 0.00018476 - layer.2.k_cache 0.00117342 0.16065517 - layer.2.v_cache 0.00000112 0.00026171 - layer.3.k_cache 0.00133382 0.17981503 - layer.3.v_cache 0.00000216 0.00043959 - layer.4.k_cache 0.00342137 0.44071367 - layer.4.v_cache 0.00000324 0.00074574 - layer.4.output 0.00018686 0.03525455 - ------------------------------------------------------------------------------------- - TOTAL 0.00263664 0.46235797 - (elements=2,523,136) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2523136 -Total Bytes 787980 -BPFP 2.4984 bits/point -EBPFP 4.9968 equivalent bits/point -MSE 0.462358 ----------------------- -------------------------------------------------------- -Time: 3.553s Load: 0.009s, Pack+Encode: 1.981s, Decode+Unpack: 1.563s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4624 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample68-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample68-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample69-layer4-item1.zst (78/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample69-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: 21,436B, BPFP=1.0667 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 65,268B, BPFP=3.2478 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,916B, BPFP=2.0858 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,724B, BPFP=3.5691 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,760B, BPFP=2.5259 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,028B, BPFP=3.5344 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,464B, BPFP=2.5111 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,692B, BPFP=3.4182 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,760B, BPFP=2.1278 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,468B, BPFP=3.5563 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 186,784B, BPFP=2.3236 -⌛️ [2/4] FRONTEND: Frontend time: 1.974s (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: 1.549s - -[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 0.02778443 4.77192134 - layer.0.v_cache 0.00000027 0.00006249 - layer.1.k_cache 0.00313050 0.33590093 - layer.1.v_cache 0.00000087 0.00021083 - layer.2.k_cache 0.00120658 0.16393754 - layer.2.v_cache 0.00000114 0.00027840 - layer.3.k_cache 0.00130588 0.17507519 - layer.3.v_cache 0.00000220 0.00046735 - layer.4.k_cache 0.00347338 0.40617837 - layer.4.v_cache 0.00000322 0.00078634 - layer.4.output 0.00013960 0.03902437 - ------------------------------------------------------------------------------------- - TOTAL 0.00267620 0.42935116 - (elements=2,250,752) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2250752 -Total Bytes 742300 -BPFP 2.6384 bits/point -EBPFP 5.2768 equivalent bits/point -MSE 0.429351 ----------------------- -------------------------------------------------------- -Time: 3.531s Load: 0.009s, Pack+Encode: 1.974s, Decode+Unpack: 1.549s ----------------------- -------------------------------------------------------- -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 0.4294 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample69-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample69-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample7-layer4-item1.zst (79/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample7-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 221, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 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, 221, 128) -Output shape: (1, 221, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.0.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.1.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.1.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.2.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.2.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.3.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.3.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.4.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.4.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.4.output: torch.Size([1, 221, 4096]) -> torch.Size([1, 1, 221, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 28,588B, BPFP=1.0106 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 88,164B, BPFP=3.1167 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 56,960B, BPFP=2.0136 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 93,088B, BPFP=3.2907 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 68,904B, BPFP=2.4358 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 95,840B, BPFP=3.3880 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 69,408B, BPFP=2.4536 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 93,624B, BPFP=3.3097 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 58,052B, BPFP=2.0522 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 95,324B, BPFP=3.3698 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 242,896B, BPFP=2.1466 -⌛️ [2/4] FRONTEND: Frontend time: 2.389s (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, 221, 128]) - layer.0.v_cache: torch.Size([1, 8, 221, 128]) - layer.1.k_cache: torch.Size([1, 8, 221, 128]) - layer.1.v_cache: torch.Size([1, 8, 221, 128]) - layer.2.k_cache: torch.Size([1, 8, 221, 128]) - layer.2.v_cache: torch.Size([1, 8, 221, 128]) - layer.3.k_cache: torch.Size([1, 8, 221, 128]) - layer.3.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.k_cache: torch.Size([1, 8, 221, 128]) - layer.4.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.output: torch.Size([1, 221, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.966s - -[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, 221, 128]) - layer.0.v_cache: torch.Size([1, 8, 221, 128]) - layer.1.k_cache: torch.Size([1, 8, 221, 128]) - layer.1.v_cache: torch.Size([1, 8, 221, 128]) - layer.2.k_cache: torch.Size([1, 8, 221, 128]) - layer.2.v_cache: torch.Size([1, 8, 221, 128]) - layer.3.k_cache: torch.Size([1, 8, 221, 128]) - layer.3.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.k_cache: torch.Size([1, 8, 221, 128]) - layer.4.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.output: torch.Size([1, 221, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02443308 4.52991579 - layer.0.v_cache 0.00000027 0.00006280 - layer.1.k_cache 0.00299519 0.32731697 - layer.1.v_cache 0.00000096 0.00018534 - layer.2.k_cache 0.00117758 0.15888779 - layer.2.v_cache 0.00000111 0.00025806 - layer.3.k_cache 0.00134364 0.17694998 - layer.3.v_cache 0.00000231 0.00046016 - layer.4.k_cache 0.00352016 0.42416033 - layer.4.v_cache 0.00000295 0.00066859 - layer.4.output 0.00021856 0.03563490 - ------------------------------------------------------------------------------------- - TOTAL 0.00245368 0.41152896 - (elements=3,168,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3168256 -Total Bytes 990848 -BPFP 2.5019 bits/point -EBPFP 5.0039 equivalent bits/point -MSE 0.411529 ----------------------- -------------------------------------------------------- -Time: 4.366s Load: 0.011s, Pack+Encode: 2.389s, Decode+Unpack: 1.966s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 221, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4115 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample7-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample7-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample70-layer4-item1.zst (80/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample70-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 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, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,692B, BPFP=1.0761 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,880B, BPFP=3.0832 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,252B, BPFP=2.1008 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,568B, BPFP=3.2961 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,292B, BPFP=2.4660 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,740B, BPFP=3.3494 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,120B, BPFP=2.4582 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,920B, BPFP=3.2667 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,964B, BPFP=2.1332 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,684B, BPFP=3.3923 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 205,068B, BPFP=2.3286 -⌛️ [2/4] FRONTEND: Frontend time: 1.952s (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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.517s - -[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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02652868 4.39545352 - layer.0.v_cache 0.00000028 0.00006145 - layer.1.k_cache 0.00309189 0.39743942 - layer.1.v_cache 0.00000082 0.00020525 - layer.2.k_cache 0.00116742 0.16173588 - layer.2.v_cache 0.00000113 0.00027281 - layer.3.k_cache 0.00134765 0.17507004 - layer.3.v_cache 0.00000211 0.00043583 - layer.4.k_cache 0.00352837 0.42719402 - layer.4.v_cache 0.00000329 0.00074714 - layer.4.output 0.00017786 0.03564754 - ------------------------------------------------------------------------------------- - TOTAL 0.00259879 0.40722897 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 791180 -BPFP 2.5669 bits/point -EBPFP 5.1338 equivalent bits/point -MSE 0.407229 ----------------------- -------------------------------------------------------- -Time: 3.478s Load: 0.009s, Pack+Encode: 1.952s, Decode+Unpack: 1.517s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4072 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample70-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample70-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample72-layer4-item1.zst (81/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample72-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 161, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 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, 161, 128) -Output shape: (1, 161, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.0.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.1.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.1.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.2.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.2.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.3.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.3.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.4.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.4.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.4.output: torch.Size([1, 161, 4096]) -> torch.Size([1, 1, 161, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,384B, BPFP=1.0377 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,576B, BPFP=3.2791 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 42,480B, BPFP=2.0613 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,856B, BPFP=3.4383 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 51,096B, BPFP=2.4794 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,404B, BPFP=3.4649 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 51,344B, BPFP=2.4915 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 69,956B, BPFP=3.3946 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 43,184B, BPFP=2.0955 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,764B, BPFP=3.4823 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 185,092B, BPFP=2.2454 -⌛️ [2/4] FRONTEND: Frontend time: 1.924s (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, 161, 128]) - layer.0.v_cache: torch.Size([1, 8, 161, 128]) - layer.1.k_cache: torch.Size([1, 8, 161, 128]) - layer.1.v_cache: torch.Size([1, 8, 161, 128]) - layer.2.k_cache: torch.Size([1, 8, 161, 128]) - layer.2.v_cache: torch.Size([1, 8, 161, 128]) - layer.3.k_cache: torch.Size([1, 8, 161, 128]) - layer.3.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.k_cache: torch.Size([1, 8, 161, 128]) - layer.4.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.output: torch.Size([1, 161, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.496s - -[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, 161, 128]) - layer.0.v_cache: torch.Size([1, 8, 161, 128]) - layer.1.k_cache: torch.Size([1, 8, 161, 128]) - layer.1.v_cache: torch.Size([1, 8, 161, 128]) - layer.2.k_cache: torch.Size([1, 8, 161, 128]) - layer.2.v_cache: torch.Size([1, 8, 161, 128]) - layer.3.k_cache: torch.Size([1, 8, 161, 128]) - layer.3.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.k_cache: torch.Size([1, 8, 161, 128]) - layer.4.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.output: torch.Size([1, 161, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02852683 4.36284074 - layer.0.v_cache 0.00000027 0.00005764 - layer.1.k_cache 0.00307027 0.35921583 - layer.1.v_cache 0.00000080 0.00019145 - layer.2.k_cache 0.00118547 0.16736600 - layer.2.v_cache 0.00000112 0.00026414 - layer.3.k_cache 0.00132345 0.18192849 - layer.3.v_cache 0.00000221 0.00043795 - layer.4.k_cache 0.00350348 0.44435845 - layer.4.v_cache 0.00000302 0.00070961 - layer.4.output 0.00018782 0.03674039 - ------------------------------------------------------------------------------------- - TOTAL 0.00274058 0.40459513 - (elements=2,308,096) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2308096 -Total Bytes 746136 -BPFP 2.5862 bits/point -EBPFP 5.1723 equivalent bits/point -MSE 0.404595 ----------------------- -------------------------------------------------------- -Time: 3.429s Load: 0.009s, Pack+Encode: 1.924s, Decode+Unpack: 1.496s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 161, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4046 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample72-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample72-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample73-layer4-item1.zst (82/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample73-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 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, 176, 128) -Output shape: (1, 176, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.output: torch.Size([1, 176, 4096]) -> torch.Size([1, 1, 176, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,624B, BPFP=1.0930 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 69,120B, BPFP=3.0682 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,728B, BPFP=2.0742 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,500B, BPFP=3.3070 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,420B, BPFP=2.4157 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,428B, BPFP=3.2594 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 55,048B, BPFP=2.4435 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,196B, BPFP=3.2047 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,588B, BPFP=2.1124 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 76,088B, BPFP=3.3775 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 214,944B, BPFP=2.3853 -⌛️ [2/4] FRONTEND: Frontend time: 1.945s (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, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.576s - -[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, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02614397 4.46205902 - layer.0.v_cache 0.00000027 0.00006124 - layer.1.k_cache 0.00303479 0.39715078 - layer.1.v_cache 0.00000100 0.00021564 - layer.2.k_cache 0.00119315 0.15964049 - layer.2.v_cache 0.00000111 0.00026116 - layer.3.k_cache 0.00127363 0.17428127 - layer.3.v_cache 0.00000250 0.00049478 - layer.4.k_cache 0.00361948 0.44995386 - layer.4.v_cache 0.00000340 0.00074903 - layer.4.output 0.00017471 0.03083915 - ------------------------------------------------------------------------------------- - TOTAL 0.00256944 0.41201599 - (elements=2,523,136) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2523136 -Total Bytes 808684 -BPFP 2.5641 bits/point -EBPFP 5.1281 equivalent bits/point -MSE 0.412016 ----------------------- -------------------------------------------------------- -Time: 3.529s Load: 0.009s, Pack+Encode: 1.945s, Decode+Unpack: 1.576s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4120 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample73-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample73-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample75-layer4-item1.zst (83/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample75-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 153, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 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, 153, 128) -Output shape: (1, 153, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.0.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.1.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.1.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.2.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.2.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.3.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.3.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.4.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.4.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.4.output: torch.Size([1, 153, 4096]) -> torch.Size([1, 1, 153, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,804B, BPFP=1.1134 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 65,944B, BPFP=3.3672 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,484B, BPFP=2.1183 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 69,288B, BPFP=3.5380 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 49,624B, BPFP=2.5339 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 70,652B, BPFP=3.6076 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,008B, BPFP=2.5535 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 67,988B, BPFP=3.4716 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 41,988B, BPFP=2.1440 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,236B, BPFP=3.6375 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 203,264B, BPFP=2.5948 -⌛️ [2/4] FRONTEND: Frontend time: 1.960s (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, 153, 128]) - layer.0.v_cache: torch.Size([1, 8, 153, 128]) - layer.1.k_cache: torch.Size([1, 8, 153, 128]) - layer.1.v_cache: torch.Size([1, 8, 153, 128]) - layer.2.k_cache: torch.Size([1, 8, 153, 128]) - layer.2.v_cache: torch.Size([1, 8, 153, 128]) - layer.3.k_cache: torch.Size([1, 8, 153, 128]) - layer.3.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.k_cache: torch.Size([1, 8, 153, 128]) - layer.4.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.output: torch.Size([1, 153, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.544s - -[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, 153, 128]) - layer.0.v_cache: torch.Size([1, 8, 153, 128]) - layer.1.k_cache: torch.Size([1, 8, 153, 128]) - layer.1.v_cache: torch.Size([1, 8, 153, 128]) - layer.2.k_cache: torch.Size([1, 8, 153, 128]) - layer.2.v_cache: torch.Size([1, 8, 153, 128]) - layer.3.k_cache: torch.Size([1, 8, 153, 128]) - layer.3.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.k_cache: torch.Size([1, 8, 153, 128]) - layer.4.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.output: torch.Size([1, 153, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02796515 4.29862069 - layer.0.v_cache 0.00000026 0.00006170 - layer.1.k_cache 0.00317772 0.34313060 - layer.1.v_cache 0.00000088 0.00021075 - layer.2.k_cache 0.00119895 0.16321616 - layer.2.v_cache 0.00000122 0.00030160 - layer.3.k_cache 0.00130332 0.17736921 - layer.3.v_cache 0.00000226 0.00049792 - layer.4.k_cache 0.00344657 0.41572491 - layer.4.v_cache 0.00000314 0.00075959 - layer.4.output 0.00015128 0.03400627 - ------------------------------------------------------------------------------------- - TOTAL 0.00269319 0.39542273 - (elements=2,193,408) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2193408 -Total Bytes 753280 -BPFP 2.7474 bits/point -EBPFP 5.4949 equivalent bits/point -MSE 0.395423 ----------------------- -------------------------------------------------------- -Time: 3.512s Load: 0.008s, Pack+Encode: 1.960s, Decode+Unpack: 1.544s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 153, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3954 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample75-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample75-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample76-layer4-item1.zst (84/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample76-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: 0.010s - ------------------------------------------------------------- -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: 21,256B, BPFP=1.0783 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,232B, BPFP=3.3600 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,560B, BPFP=2.1084 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,192B, BPFP=3.5609 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,872B, BPFP=2.5808 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,004B, BPFP=3.6021 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,376B, BPFP=2.5556 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 69,160B, BPFP=3.5085 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,396B, BPFP=2.1508 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,884B, BPFP=3.6467 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 191,964B, BPFP=2.4346 -⌛️ [2/4] FRONTEND: Frontend time: 1.990s (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: 1.539s - -[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 0.02770888 4.53683392 - layer.0.v_cache 0.00000027 0.00005943 - layer.1.k_cache 0.00299767 0.32105602 - layer.1.v_cache 0.00000089 0.00020584 - layer.2.k_cache 0.00118955 0.16443228 - layer.2.v_cache 0.00000124 0.00028884 - layer.3.k_cache 0.00128217 0.17854852 - layer.3.v_cache 0.00000233 0.00048510 - layer.4.k_cache 0.00329405 0.39207800 - layer.4.v_cache 0.00000340 0.00081004 - layer.4.output 0.00014754 0.03448170 - ------------------------------------------------------------------------------------- - TOTAL 0.00264790 0.40948034 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 746896 -BPFP 2.7065 bits/point -EBPFP 5.4129 equivalent bits/point -MSE 0.409480 ----------------------- -------------------------------------------------------- -Time: 3.538s Load: 0.010s, Pack+Encode: 1.990s, Decode+Unpack: 1.539s ----------------------- -------------------------------------------------------- -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 0.4095 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample76-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample76-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample78-layer4-item1.zst (85/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample78-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 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, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,380B, BPFP=1.0808 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,308B, BPFP=3.1577 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,304B, BPFP=2.0943 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,736B, BPFP=3.3624 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,628B, BPFP=2.4791 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 75,008B, BPFP=3.4675 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,280B, BPFP=2.5092 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,008B, BPFP=3.3288 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,364B, BPFP=2.1433 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,812B, BPFP=3.4122 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 234,148B, BPFP=2.7060 -⌛️ [2/4] FRONTEND: Frontend time: 1.964s (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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.557s - -[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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02814460 4.71755349 - layer.0.v_cache 0.00000027 0.00005931 - layer.1.k_cache 0.00305883 0.37714465 - layer.1.v_cache 0.00000085 0.00020723 - layer.2.k_cache 0.00120620 0.17088188 - layer.2.v_cache 0.00000155 0.00030124 - layer.3.k_cache 0.00127672 0.17789212 - layer.3.v_cache 0.00000231 0.00047471 - layer.4.k_cache 0.00421318 0.46916610 - layer.4.v_cache 0.00000320 0.00071679 - layer.4.output 0.00020059 0.03255089 - ------------------------------------------------------------------------------------- - TOTAL 0.00276501 0.43175722 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 818976 -BPFP 2.7042 bits/point -EBPFP 5.4085 equivalent bits/point -MSE 0.431757 ----------------------- -------------------------------------------------------- -Time: 3.530s Load: 0.010s, Pack+Encode: 1.964s, Decode+Unpack: 1.557s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4318 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample78-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample78-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample79-layer4-item1.zst (86/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample79-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 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, 180, 128) -Output shape: (1, 180, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.output: torch.Size([1, 180, 4096]) -> torch.Size([1, 1, 180, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,832B, BPFP=1.0344 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,756B, BPFP=2.9842 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,972B, BPFP=1.9953 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 74,940B, BPFP=3.2526 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 55,144B, BPFP=2.3934 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,992B, BPFP=3.2549 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,768B, BPFP=2.3771 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,952B, BPFP=3.1663 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,764B, BPFP=2.0297 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,760B, BPFP=3.2014 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 214,140B, BPFP=2.3236 -⌛️ [2/4] FRONTEND: Frontend time: 1.978s (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, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.603s - -[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, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02888792 4.55775418 - layer.0.v_cache 0.00000029 0.00006110 - layer.1.k_cache 0.00301893 0.38521894 - layer.1.v_cache 0.00000086 0.00019649 - layer.2.k_cache 0.00125228 0.16280028 - layer.2.v_cache 0.00000118 0.00026638 - layer.3.k_cache 0.00126354 0.16927043 - layer.3.v_cache 0.00000246 0.00048256 - layer.4.k_cache 0.00353636 0.39392331 - layer.4.v_cache 0.00000320 0.00070287 - layer.4.output 0.00017167 0.03495494 - ------------------------------------------------------------------------------------- - TOTAL 0.00276098 0.41503545 - (elements=2,580,480) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2580480 -Total Bytes 806020 -BPFP 2.4988 bits/point -EBPFP 4.9976 equivalent bits/point -MSE 0.415035 ----------------------- -------------------------------------------------------- -Time: 3.590s Load: 0.009s, Pack+Encode: 1.978s, Decode+Unpack: 1.603s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4150 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample79-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample79-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample8-layer4-item1.zst (87/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample8-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 212, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 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, 212, 128) -Output shape: (1, 212, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.0.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.1.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.1.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.2.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.2.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.3.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.3.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.4.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.4.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.4.output: torch.Size([1, 212, 4096]) -> torch.Size([1, 1, 212, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 28,860B, BPFP=1.0635 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 86,604B, BPFP=3.1915 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 55,472B, BPFP=2.0442 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 91,728B, BPFP=3.3803 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 67,632B, BPFP=2.4923 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 93,380B, BPFP=3.4412 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 68,272B, BPFP=2.5159 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 90,828B, BPFP=3.3471 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 55,300B, BPFP=2.0379 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 93,344B, BPFP=3.4399 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 253,456B, BPFP=2.3351 -⌛️ [2/4] FRONTEND: Frontend time: 2.386s (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, 212, 128]) - layer.0.v_cache: torch.Size([1, 8, 212, 128]) - layer.1.k_cache: torch.Size([1, 8, 212, 128]) - layer.1.v_cache: torch.Size([1, 8, 212, 128]) - layer.2.k_cache: torch.Size([1, 8, 212, 128]) - layer.2.v_cache: torch.Size([1, 8, 212, 128]) - layer.3.k_cache: torch.Size([1, 8, 212, 128]) - layer.3.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.k_cache: torch.Size([1, 8, 212, 128]) - layer.4.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.output: torch.Size([1, 212, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.989s - -[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, 212, 128]) - layer.0.v_cache: torch.Size([1, 8, 212, 128]) - layer.1.k_cache: torch.Size([1, 8, 212, 128]) - layer.1.v_cache: torch.Size([1, 8, 212, 128]) - layer.2.k_cache: torch.Size([1, 8, 212, 128]) - layer.2.v_cache: torch.Size([1, 8, 212, 128]) - layer.3.k_cache: torch.Size([1, 8, 212, 128]) - layer.3.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.k_cache: torch.Size([1, 8, 212, 128]) - layer.4.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.output: torch.Size([1, 212, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02614697 3.80197431 - layer.0.v_cache 0.00000027 0.00005671 - layer.1.k_cache 0.00297715 0.33296384 - layer.1.v_cache 0.00000079 0.00018830 - layer.2.k_cache 0.00116049 0.16089191 - layer.2.v_cache 0.00000116 0.00027385 - layer.3.k_cache 0.00132944 0.18015620 - layer.3.v_cache 0.00000234 0.00045577 - layer.4.k_cache 0.00407070 0.39250978 - layer.4.v_cache 0.00000311 0.00068019 - layer.4.output 0.00021146 0.03292895 - ------------------------------------------------------------------------------------- - TOTAL 0.00260988 0.35727619 - (elements=3,039,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3039232 -Total Bytes 984876 -BPFP 2.5924 bits/point -EBPFP 5.1849 equivalent bits/point -MSE 0.357276 ----------------------- -------------------------------------------------------- -Time: 4.386s Load: 0.011s, Pack+Encode: 2.386s, Decode+Unpack: 1.989s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 212, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3573 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample8-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample8-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample80-layer4-item1.zst (88/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample80-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 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, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 23,836B, BPFP=1.0827 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,188B, BPFP=3.0972 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,460B, BPFP=2.1103 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,984B, BPFP=3.2696 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,196B, BPFP=2.4617 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,308B, BPFP=3.3298 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,992B, BPFP=2.4524 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,964B, BPFP=3.2233 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,680B, BPFP=2.1203 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,472B, BPFP=3.2918 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 199,036B, BPFP=2.2601 -⌛️ [2/4] FRONTEND: Frontend time: 1.978s (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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.572s - -[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, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02862963 4.58216751 - layer.0.v_cache 0.00000027 0.00005863 - layer.1.k_cache 0.00312686 0.40670759 - layer.1.v_cache 0.00000083 0.00020510 - layer.2.k_cache 0.00115513 0.15956105 - layer.2.v_cache 0.00000117 0.00027308 - layer.3.k_cache 0.00132852 0.17446897 - layer.3.v_cache 0.00000218 0.00043305 - layer.4.k_cache 0.00345250 0.40647658 - layer.4.v_cache 0.00000308 0.00070573 - layer.4.output 0.00017775 0.03258612 - ------------------------------------------------------------------------------------- - TOTAL 0.00274366 0.41867156 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 781116 -BPFP 2.5342 bits/point -EBPFP 5.0685 equivalent bits/point -MSE 0.418672 ----------------------- -------------------------------------------------------- -Time: 3.560s Load: 0.010s, Pack+Encode: 1.978s, Decode+Unpack: 1.572s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4187 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample80-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample80-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample81-layer4-item1.zst (89/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample81-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 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, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,352B, BPFP=1.0558 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,984B, BPFP=3.3121 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,420B, BPFP=2.0481 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,360B, BPFP=3.4790 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,848B, BPFP=2.5142 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 70,988B, BPFP=3.5101 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,920B, BPFP=2.5178 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 69,556B, BPFP=3.4393 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,964B, BPFP=2.1244 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,484B, BPFP=3.5346 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 179,016B, BPFP=2.2129 -⌛️ [2/4] FRONTEND: Frontend time: 1.989s (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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.583s - -[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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02915672 4.35853519 - layer.0.v_cache 0.00000028 0.00006025 - layer.1.k_cache 0.00318231 0.32021911 - layer.1.v_cache 0.00000085 0.00020358 - layer.2.k_cache 0.00121414 0.16588727 - layer.2.v_cache 0.00000113 0.00027541 - layer.3.k_cache 0.00130024 0.17890080 - layer.3.v_cache 0.00000226 0.00046368 - layer.4.k_cache 0.00343194 0.41425222 - layer.4.v_cache 0.00000318 0.00074848 - layer.4.output 0.00014314 0.03301661 - ------------------------------------------------------------------------------------- - TOTAL 0.00277611 0.39797232 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 735892 -BPFP 2.5991 bits/point -EBPFP 5.1982 equivalent bits/point -MSE 0.397972 ----------------------- -------------------------------------------------------- -Time: 3.581s Load: 0.008s, Pack+Encode: 1.989s, Decode+Unpack: 1.583s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3980 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample81-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample81-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample88-layer4-item1.zst (90/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample88-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 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, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,040B, BPFP=1.0189 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,536B, BPFP=3.1220 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,148B, BPFP=2.0871 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,644B, BPFP=3.3119 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 53,756B, BPFP=2.4850 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 72,920B, BPFP=3.3709 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,092B, BPFP=2.5006 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,076B, BPFP=3.3319 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,408B, BPFP=2.1453 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,972B, BPFP=3.3733 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 196,528B, BPFP=2.2713 -⌛️ [2/4] FRONTEND: Frontend time: 1.950s (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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.564s - -[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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02741317 4.19394264 - layer.0.v_cache 0.00000027 0.00005826 - layer.1.k_cache 0.00313704 0.33845890 - layer.1.v_cache 0.00000078 0.00018703 - layer.2.k_cache 0.00120918 0.16851624 - layer.2.v_cache 0.00000116 0.00027004 - layer.3.k_cache 0.00136700 0.18770537 - layer.3.v_cache 0.00000220 0.00044559 - layer.4.k_cache 0.00352943 0.47095918 - layer.4.v_cache 0.00000294 0.00068468 - layer.4.output 0.00020829 0.03872074 - ------------------------------------------------------------------------------------- - TOTAL 0.00267831 0.39400792 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 775120 -BPFP 2.5594 bits/point -EBPFP 5.1189 equivalent bits/point -MSE 0.394008 ----------------------- -------------------------------------------------------- -Time: 3.522s Load: 0.009s, Pack+Encode: 1.950s, Decode+Unpack: 1.564s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3940 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample88-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample88-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample89-layer4-item1.zst (91/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample89-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 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, 162, 128) -Output shape: (1, 162, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.output: torch.Size([1, 162, 4096]) -> torch.Size([1, 1, 162, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,120B, BPFP=1.0185 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,380B, BPFP=3.2494 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 43,940B, BPFP=2.1190 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,756B, BPFP=3.4122 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 51,892B, BPFP=2.5025 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,600B, BPFP=3.4529 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 51,948B, BPFP=2.5052 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,076B, BPFP=3.3794 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 44,428B, BPFP=2.1426 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,088B, BPFP=3.5247 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 189,052B, BPFP=2.2793 -⌛️ [2/4] FRONTEND: Frontend time: 2.019s (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, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.530s - -[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, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02771055 4.36601502 - layer.0.v_cache 0.00000026 0.00006704 - layer.1.k_cache 0.00323549 0.38147566 - layer.1.v_cache 0.00000081 0.00021430 - layer.2.k_cache 0.00117871 0.16862302 - layer.2.v_cache 0.00000115 0.00029176 - layer.3.k_cache 0.00135828 0.19396678 - layer.3.v_cache 0.00000214 0.00047524 - layer.4.k_cache 0.00384025 0.51772572 - layer.4.v_cache 0.00000294 0.00071397 - layer.4.output 0.00023325 0.04122769 - ------------------------------------------------------------------------------------- - TOTAL 0.00273311 0.41389138 - (elements=2,322,432) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2322432 -Total Bytes 755280 -BPFP 2.6017 bits/point -EBPFP 5.2034 equivalent bits/point -MSE 0.413891 ----------------------- -------------------------------------------------------- -Time: 3.558s Load: 0.009s, Pack+Encode: 2.019s, Decode+Unpack: 1.530s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4139 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample89-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample89-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample9-layer4-item1.zst (92/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample9-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 193, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 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, 193, 128) -Output shape: (1, 193, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.0.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.1.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.1.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.2.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.2.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.3.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.3.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.4.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.4.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.4.output: torch.Size([1, 193, 4096]) -> torch.Size([1, 1, 193, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 25,920B, BPFP=1.0492 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 80,588B, BPFP=3.2621 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 49,392B, BPFP=1.9994 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 89,552B, BPFP=3.6250 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 62,280B, BPFP=2.5210 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 89,184B, BPFP=3.6101 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 62,980B, BPFP=2.5494 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 87,320B, BPFP=3.5347 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 50,748B, BPFP=2.0542 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 91,212B, BPFP=3.6922 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 235,576B, BPFP=2.3840 -⌛️ [2/4] FRONTEND: Frontend time: 2.371s (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, 193, 128]) - layer.0.v_cache: torch.Size([1, 8, 193, 128]) - layer.1.k_cache: torch.Size([1, 8, 193, 128]) - layer.1.v_cache: torch.Size([1, 8, 193, 128]) - layer.2.k_cache: torch.Size([1, 8, 193, 128]) - layer.2.v_cache: torch.Size([1, 8, 193, 128]) - layer.3.k_cache: torch.Size([1, 8, 193, 128]) - layer.3.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.k_cache: torch.Size([1, 8, 193, 128]) - layer.4.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.output: torch.Size([1, 193, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.966s - -[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, 193, 128]) - layer.0.v_cache: torch.Size([1, 8, 193, 128]) - layer.1.k_cache: torch.Size([1, 8, 193, 128]) - layer.1.v_cache: torch.Size([1, 8, 193, 128]) - layer.2.k_cache: torch.Size([1, 8, 193, 128]) - layer.2.v_cache: torch.Size([1, 8, 193, 128]) - layer.3.k_cache: torch.Size([1, 8, 193, 128]) - layer.3.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.k_cache: torch.Size([1, 8, 193, 128]) - layer.4.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.output: torch.Size([1, 193, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02858143 4.24366246 - layer.0.v_cache 0.00000027 0.00006283 - layer.1.k_cache 0.00303266 0.33575550 - layer.1.v_cache 0.00000097 0.00021570 - layer.2.k_cache 0.00124314 0.15915453 - layer.2.v_cache 0.00000124 0.00030043 - layer.3.k_cache 0.00125873 0.16871764 - layer.3.v_cache 0.00000241 0.00049155 - layer.4.k_cache 0.00343145 0.35769823 - layer.4.v_cache 0.00000350 0.00083191 - layer.4.output 0.00017011 0.03128944 - ------------------------------------------------------------------------------------- - TOTAL 0.00273116 0.38514632 - (elements=2,766,848) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2766848 -Total Bytes 924752 -BPFP 2.6738 bits/point -EBPFP 5.3476 equivalent bits/point -MSE 0.385146 ----------------------- -------------------------------------------------------- -Time: 4.348s Load: 0.011s, Pack+Encode: 2.371s, Decode+Unpack: 1.966s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 193, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3851 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample9-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample9-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample90-layer4-item1.zst (93/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample90-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 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, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 26,336B, BPFP=1.0886 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 70,996B, BPFP=2.9347 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,596B, BPFP=1.9261 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 75,624B, BPFP=3.1260 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 55,220B, BPFP=2.2826 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 76,428B, BPFP=3.1592 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 55,032B, BPFP=2.2748 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 72,316B, BPFP=2.9893 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,304B, BPFP=1.9554 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 78,180B, BPFP=3.2316 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 219,832B, BPFP=2.2717 -⌛️ [2/4] FRONTEND: Frontend time: 2.021s (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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.597s - -[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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02726509 4.53452911 - layer.0.v_cache 0.00000027 0.00006386 - layer.1.k_cache 0.00303300 0.42280526 - layer.1.v_cache 0.00000085 0.00020713 - layer.2.k_cache 0.00116837 0.15301258 - layer.2.v_cache 0.00000118 0.00027676 - layer.3.k_cache 0.00127025 0.16912364 - layer.3.v_cache 0.00000234 0.00046384 - layer.4.k_cache 0.00356333 0.40704979 - layer.4.v_cache 0.00000334 0.00073277 - layer.4.output 0.00017162 0.03294722 - ------------------------------------------------------------------------------------- - TOTAL 0.00264246 0.41571812 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 823864 -BPFP 2.4325 bits/point -EBPFP 4.8650 equivalent bits/point -MSE 0.415718 ----------------------- -------------------------------------------------------- -Time: 3.628s Load: 0.010s, Pack+Encode: 2.021s, Decode+Unpack: 1.597s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4157 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample90-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample90-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample91-layer4-item1.zst (94/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample91-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 26,600B, BPFP=1.0995 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 70,236B, BPFP=2.9033 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,896B, BPFP=1.9385 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 77,000B, BPFP=3.1829 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 55,100B, BPFP=2.2776 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 80,060B, BPFP=3.3094 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 55,088B, BPFP=2.2771 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,840B, BPFP=2.9696 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,280B, BPFP=1.9544 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 77,852B, BPFP=3.2181 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 217,888B, BPFP=2.2517 -⌛️ [2/4] FRONTEND: Frontend time: 1.972s (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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.569s - -[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, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02698849 5.14887799 - layer.0.v_cache 0.00000027 0.00006330 - layer.1.k_cache 0.00302237 0.41685571 - layer.1.v_cache 0.00000087 0.00021127 - layer.2.k_cache 0.00117769 0.15356951 - layer.2.v_cache 0.00000119 0.00027441 - layer.3.k_cache 0.00126503 0.16674005 - layer.3.v_cache 0.00000236 0.00046589 - layer.4.k_cache 0.00350408 0.40530702 - layer.4.v_cache 0.00000327 0.00073471 - layer.4.output 0.00016083 0.03014655 - ------------------------------------------------------------------------------------- - TOTAL 0.00261492 0.45812043 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 825840 -BPFP 2.4384 bits/point -EBPFP 4.8767 equivalent bits/point -MSE 0.458120 ----------------------- -------------------------------------------------------- -Time: 3.553s Load: 0.012s, Pack+Encode: 1.972s, Decode+Unpack: 1.569s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4581 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample91-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample91-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample93-layer4-item1.zst (95/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample93-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 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, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 21,116B, BPFP=1.0441 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 66,256B, BPFP=3.2761 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 41,200B, BPFP=2.0372 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,004B, BPFP=3.4614 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 50,784B, BPFP=2.5111 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,356B, BPFP=3.5283 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 50,880B, BPFP=2.5158 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 68,920B, BPFP=3.4078 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 42,540B, BPFP=2.1034 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 71,984B, BPFP=3.5593 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 183,400B, BPFP=2.2671 -⌛️ [2/4] FRONTEND: Frontend time: 1.987s (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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.592s - -[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, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02813752 4.58550678 - layer.0.v_cache 0.00000027 0.00006579 - layer.1.k_cache 0.00306488 0.32871596 - layer.1.v_cache 0.00000089 0.00020842 - layer.2.k_cache 0.00118156 0.16866401 - layer.2.v_cache 0.00000118 0.00030024 - layer.3.k_cache 0.00130882 0.18148641 - layer.3.v_cache 0.00000213 0.00045795 - layer.4.k_cache 0.00347855 0.42228892 - layer.4.v_cache 0.00000307 0.00075208 - layer.4.output 0.00015371 0.03046434 - ------------------------------------------------------------------------------------- - TOTAL 0.00269955 0.41502171 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 738440 -BPFP 2.6081 bits/point -EBPFP 5.2162 equivalent bits/point -MSE 0.415022 ----------------------- -------------------------------------------------------- -Time: 3.589s Load: 0.010s, Pack+Encode: 1.987s, Decode+Unpack: 1.592s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4150 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample93-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample93-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample94-layer4-item1.zst (96/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample94-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 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, 177, 128) -Output shape: (1, 177, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.output: torch.Size([1, 177, 4096]) -> torch.Size([1, 1, 177, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 24,332B, BPFP=1.0740 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,368B, BPFP=2.9735 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,848B, BPFP=2.0678 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,264B, BPFP=3.1896 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,748B, BPFP=2.4165 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,756B, BPFP=3.2996 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,432B, BPFP=2.4025 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,612B, BPFP=3.1608 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,688B, BPFP=2.0607 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,968B, BPFP=3.2648 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 209,076B, BPFP=2.3071 -⌛️ [2/4] FRONTEND: Frontend time: 2.012s (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, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.526s - -[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, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02696669 4.18276719 - layer.0.v_cache 0.00000028 0.00005815 - layer.1.k_cache 0.00303349 0.41583131 - layer.1.v_cache 0.00000089 0.00019906 - layer.2.k_cache 0.00118683 0.16811094 - layer.2.v_cache 0.00000112 0.00027320 - layer.3.k_cache 0.00129356 0.17432455 - layer.3.v_cache 0.00000212 0.00043405 - layer.4.k_cache 0.00349999 0.44642575 - layer.4.v_cache 0.00000320 0.00071476 - layer.4.output 0.00017955 0.02655349 - ------------------------------------------------------------------------------------- - TOTAL 0.00262188 0.39252521 - (elements=2,537,472) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2537472 -Total Bytes 796092 -BPFP 2.5099 bits/point -EBPFP 5.0197 equivalent bits/point -MSE 0.392525 ----------------------- -------------------------------------------------------- -Time: 3.547s Load: 0.010s, Pack+Encode: 2.012s, Decode+Unpack: 1.526s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.3925 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample94-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample94-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample95-layer4-item1.zst (97/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample95-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 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, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,448B, BPFP=1.0377 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,260B, BPFP=3.1093 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 45,332B, BPFP=2.0956 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 72,960B, BPFP=3.3728 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 52,716B, BPFP=2.4369 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,092B, BPFP=3.4251 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 53,624B, BPFP=2.4789 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,612B, BPFP=3.2642 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,096B, BPFP=2.1309 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 74,192B, BPFP=3.4297 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 192,520B, BPFP=2.2249 -⌛️ [2/4] FRONTEND: Frontend time: 2.004s (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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.576s - -[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, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02729142 4.45269342 - layer.0.v_cache 0.00000026 0.00006031 - layer.1.k_cache 0.00319685 0.39886055 - layer.1.v_cache 0.00000078 0.00019962 - layer.2.k_cache 0.00117458 0.16589933 - layer.2.v_cache 0.00000113 0.00028175 - layer.3.k_cache 0.00133789 0.18646335 - layer.3.v_cache 0.00000210 0.00046413 - layer.4.k_cache 0.00349792 0.49635685 - layer.4.v_cache 0.00000290 0.00071432 - layer.4.output 0.00017352 0.02505480 - ------------------------------------------------------------------------------------- - TOTAL 0.00265714 0.41444378 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 771852 -BPFP 2.5486 bits/point -EBPFP 5.0973 equivalent bits/point -MSE 0.414444 ----------------------- -------------------------------------------------------- -Time: 3.592s Load: 0.011s, Pack+Encode: 2.004s, Decode+Unpack: 1.576s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4144 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample95-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample95-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample96-layer4-item1.zst (98/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample96-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 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, 162, 128) -Output shape: (1, 162, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.output: torch.Size([1, 162, 4096]) -> torch.Size([1, 1, 162, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 22,068B, BPFP=1.0642 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,716B, BPFP=3.2656 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 43,780B, BPFP=2.1113 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 70,936B, BPFP=3.4209 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 52,388B, BPFP=2.5264 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 71,708B, BPFP=3.4581 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 52,172B, BPFP=2.5160 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,444B, BPFP=3.3972 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 44,616B, BPFP=2.1516 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 72,720B, BPFP=3.5069 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 189,256B, BPFP=2.2817 -⌛️ [2/4] FRONTEND: Frontend time: 1.975s (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, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.561s - -[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, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02640657 4.37847524 - layer.0.v_cache 0.00000026 0.00006181 - layer.1.k_cache 0.00309680 0.38867633 - layer.1.v_cache 0.00000084 0.00020886 - layer.2.k_cache 0.00116892 0.16535100 - layer.2.v_cache 0.00000121 0.00029797 - layer.3.k_cache 0.00130413 0.17890290 - layer.3.v_cache 0.00000229 0.00046748 - layer.4.k_cache 0.00341342 0.45868118 - layer.4.v_cache 0.00000318 0.00078057 - layer.4.output 0.00016848 0.03542742 - ------------------------------------------------------------------------------------- - TOTAL 0.00257654 0.40811521 - (elements=2,322,432) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2322432 -Total Bytes 757804 -BPFP 2.6104 bits/point -EBPFP 5.2208 equivalent bits/point -MSE 0.408115 ----------------------- -------------------------------------------------------- -Time: 3.544s Load: 0.009s, Pack+Encode: 1.975s, Decode+Unpack: 1.561s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4081 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample96-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample96-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample98-layer4-item1.zst (99/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample98-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: 0.012s - ------------------------------------------------------------- -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: 22,564B, BPFP=1.0190 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 67,908B, BPFP=3.0667 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,044B, BPFP=2.0793 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 76,300B, BPFP=3.4456 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,016B, BPFP=2.4393 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 73,948B, BPFP=3.3394 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,476B, BPFP=2.4601 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 71,980B, BPFP=3.2505 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 46,468B, BPFP=2.0984 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,340B, BPFP=3.3120 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 202,480B, BPFP=2.2859 -⌛️ [2/4] FRONTEND: Frontend time: 2.015s (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: 1.552s - -[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 0.02785747 4.21743404 - layer.0.v_cache 0.00000027 0.00006324 - layer.1.k_cache 0.00305214 0.38922031 - layer.1.v_cache 0.00000084 0.00021459 - layer.2.k_cache 0.00114171 0.16272760 - layer.2.v_cache 0.00000110 0.00027686 - layer.3.k_cache 0.00128082 0.18106915 - layer.3.v_cache 0.00000218 0.00044786 - layer.4.k_cache 0.00367794 0.46966491 - layer.4.v_cache 0.00000299 0.00070945 - layer.4.output 0.00016652 0.03184443 - ------------------------------------------------------------------------------------- - TOTAL 0.00269168 0.39637184 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 789524 -BPFP 2.5467 bits/point -EBPFP 5.0934 equivalent bits/point -MSE 0.396372 ----------------------- -------------------------------------------------------- -Time: 3.579s Load: 0.012s, Pack+Encode: 2.015s, Decode+Unpack: 1.552s ----------------------- -------------------------------------------------------- -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 0.3964 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample98-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample98-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample99-layer4-item1.zst (100/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample99-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 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, 177, 128) -Output shape: (1, 177, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.output: torch.Size([1, 177, 4096]) -> torch.Size([1, 1, 177, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 26,996B, BPFP=1.1916 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 68,024B, BPFP=3.0025 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 46,184B, BPFP=2.0385 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 71,424B, BPFP=3.1525 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 54,224B, BPFP=2.3934 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 74,380B, BPFP=3.2830 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 54,424B, BPFP=2.4022 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 70,940B, BPFP=3.1312 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 47,112B, BPFP=2.0794 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 73,948B, BPFP=3.2639 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 201,268B, BPFP=2.2209 -⌛️ [2/4] FRONTEND: Frontend time: 1.979s (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, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.582s - -[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, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02698118 5.30993652 - layer.0.v_cache 0.00000026 0.00006040 - layer.1.k_cache 0.00304540 0.37583501 - layer.1.v_cache 0.00000079 0.00019308 - layer.2.k_cache 0.00119876 0.16055544 - layer.2.v_cache 0.00000114 0.00027279 - layer.3.k_cache 0.00131734 0.17785084 - layer.3.v_cache 0.00000220 0.00045178 - layer.4.k_cache 0.00342134 0.43782931 - layer.4.v_cache 0.00000318 0.00076732 - layer.4.output 0.00016969 0.02989363 - ------------------------------------------------------------------------------------- - TOTAL 0.00261788 0.47023764 - (elements=2,537,472) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2537472 -Total Bytes 788924 -BPFP 2.4873 bits/point -EBPFP 4.9746 equivalent bits/point -MSE 0.470238 ----------------------- -------------------------------------------------------- -Time: 3.572s Load: 0.011s, Pack+Encode: 1.979s, Decode+Unpack: 1.582s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4702 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample99-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample99-layer4-item1.zst ------------------------- ---------------------------- -TOTAL PROCESSING SUMMARY ------------------------- ---------------------------- -Total files 100 -Avg BPFP 2.5742 bits/point -Avg EBPFP 5.1485 equivalent bits/point -Avg MSE 0.410715 -Avg Time 4.073s ------------------------- ---------------------------- +version https://git-lfs.github.com/spec/v1 +oid sha256:a0c81503251f99ccb6605aa194816af1094aec09adb9d3c3a51143173a302abb +size 1125774 diff --git a/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log index 04c33e06ab6ba3e6a1ffdd94dc924ce3530a210e..cd18d46fc3191011b0f9d24fb18642cfcaa16062 100644 --- a/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log +++ b/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log @@ -1,16958 +1,3 @@ -Experiment: dtufc_elic-featurecoding_qwen_individual -Log file: output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log -DTUFCCodecConfig: - arch: elic-featurecoding - handler: qwen - checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_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/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar -Checkpoint epoch: 520 -Loaded elic-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 elic-featurecoding -Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_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_winogrande -Output output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande ----------------- ------------------------------------------------------------------------------------------------------------------- -Files found: 100 ----------------------------------------------------------------------- - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample0-layer4-item1.zst (1/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample0-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: 0.005s - ------------------------------------------------------------- -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: 14,160B, BPFP=1.1062 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 45,088B, BPFP=3.5225 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 30,044B, BPFP=2.3472 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,760B, BPFP=3.6531 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,628B, BPFP=2.7053 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 47,336B, BPFP=3.6981 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 35,256B, BPFP=2.7544 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 46,688B, BPFP=3.6475 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 30,700B, BPFP=2.3984 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,644B, BPFP=3.7222 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 124,932B, BPFP=2.4401 -⌛️ [2/4] FRONTEND: Frontend time: 2.223s (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: 1.347s - -[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 0.03019863 4.46799896 - layer.0.v_cache 0.00000027 0.00006238 - layer.1.k_cache 0.00347471 0.39595768 - layer.1.v_cache 0.00000091 0.00019613 - layer.2.k_cache 0.00113692 0.17120970 - layer.2.v_cache 0.00000108 0.00027495 - layer.3.k_cache 0.00132974 0.19269245 - layer.3.v_cache 0.00000211 0.00046966 - layer.4.k_cache 0.00328898 0.48658508 - layer.4.v_cache 0.00000306 0.00072893 - layer.4.output 0.00017021 0.04152079 - ------------------------------------------------------------------------------------- - TOTAL 0.00286552 0.42016136 - (elements=1,433,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1433600 -Total Bytes 503236 -BPFP 2.8082 bits/point -EBPFP 5.6165 equivalent bits/point -MSE 0.420161 ----------------------- -------------------------------------------------------- -Time: 3.575s Load: 0.005s, Pack+Encode: 2.223s, Decode+Unpack: 1.347s ----------------------- -------------------------------------------------------- -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 0.4202 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample0-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample0-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample1-layer4-item1.zst (2/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample1-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: 0.008s - ------------------------------------------------------------- -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: 13,644B, BPFP=1.0877 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 44,928B, BPFP=3.5816 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,144B, BPFP=2.2436 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,620B, BPFP=3.7165 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,100B, BPFP=2.7184 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 47,320B, BPFP=3.7723 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,012B, BPFP=2.7114 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 46,336B, BPFP=3.6939 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,640B, BPFP=2.2832 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,704B, BPFP=3.8029 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 117,344B, BPFP=2.3386 -⌛️ [2/4] FRONTEND: Frontend time: 1.814s (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: 1.352s - -[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 0.03033472 4.27792794 - layer.0.v_cache 0.00000026 0.00005879 - layer.1.k_cache 0.00340990 0.35974806 - layer.1.v_cache 0.00000094 0.00019739 - layer.2.k_cache 0.00111132 0.17037166 - layer.2.v_cache 0.00000109 0.00027899 - layer.3.k_cache 0.00132423 0.19658124 - layer.3.v_cache 0.00000205 0.00046103 - layer.4.k_cache 0.00330658 0.46409852 - layer.4.v_cache 0.00000307 0.00076867 - layer.4.output 0.00020590 0.04946524 - ------------------------------------------------------------------------------------- - TOTAL 0.00287984 0.40488238 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 488792 -BPFP 2.7833 bits/point -EBPFP 5.5666 equivalent bits/point -MSE 0.404882 ----------------------- -------------------------------------------------------- -Time: 3.173s Load: 0.008s, Pack+Encode: 1.814s, Decode+Unpack: 1.352s ----------------------- -------------------------------------------------------- -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 0.4049 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample1-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample1-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample10-layer4-item1.zst (3/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample10-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: 0.008s - ------------------------------------------------------------- -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: 13,140B, BPFP=1.0921 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,936B, BPFP=3.4854 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,296B, BPFP=2.1024 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,892B, BPFP=3.7311 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,760B, BPFP=2.6396 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,344B, BPFP=3.7686 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,596B, BPFP=2.7091 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,828B, BPFP=3.7257 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,408B, BPFP=2.1117 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,392B, BPFP=3.7726 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 112,900B, BPFP=2.3458 -⌛️ [2/4] FRONTEND: Frontend time: 1.758s (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: 1.351s - -[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 0.03043811 4.09780624 - layer.0.v_cache 0.00000028 0.00006155 - layer.1.k_cache 0.00344812 0.34431750 - layer.1.v_cache 0.00000082 0.00020819 - layer.2.k_cache 0.00115027 0.17144088 - layer.2.v_cache 0.00000109 0.00027816 - layer.3.k_cache 0.00130609 0.18907541 - layer.3.v_cache 0.00000216 0.00047509 - layer.4.k_cache 0.00321842 0.40787774 - layer.4.v_cache 0.00000315 0.00075800 - layer.4.output 0.00017058 0.04457426 - ------------------------------------------------------------------------------------- - TOTAL 0.00287506 0.38504256 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 463492 -BPFP 2.7515 bits/point -EBPFP 5.5031 equivalent bits/point -MSE 0.385043 ----------------------- -------------------------------------------------------- -Time: 3.116s Load: 0.008s, Pack+Encode: 1.758s, Decode+Unpack: 1.351s ----------------------- -------------------------------------------------------- -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 0.3850 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample10-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample10-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample100-layer4-item1.zst (4/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample100-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 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, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,712B, BPFP=1.1286 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,388B, BPFP=3.5856 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,484B, BPFP=2.0849 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,204B, BPFP=3.7468 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,804B, BPFP=2.5572 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,168B, BPFP=3.8324 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,480B, BPFP=2.7060 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,968B, BPFP=3.7259 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,120B, BPFP=2.0526 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,312B, BPFP=3.8452 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,488B, BPFP=2.3413 -⌛️ [2/4] FRONTEND: Frontend time: 1.763s (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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.353s - -[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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03339012 4.45748589 - layer.0.v_cache 0.00000028 0.00006246 - layer.1.k_cache 0.00350260 0.33151145 - layer.1.v_cache 0.00000079 0.00020477 - layer.2.k_cache 0.00114403 0.16598305 - layer.2.v_cache 0.00000109 0.00028233 - layer.3.k_cache 0.00131544 0.19898126 - layer.3.v_cache 0.00000207 0.00048717 - layer.4.k_cache 0.00326494 0.38235773 - layer.4.v_cache 0.00000300 0.00077788 - layer.4.output 0.00017308 0.03899236 - ------------------------------------------------------------------------------------- - TOTAL 0.00309405 0.40672167 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 435128 -BPFP 2.7593 bits/point -EBPFP 5.5186 equivalent bits/point -MSE 0.406722 ----------------------- -------------------------------------------------------- -Time: 3.123s Load: 0.007s, Pack+Encode: 1.763s, Decode+Unpack: 1.353s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4067 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample100-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample100-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample103-layer4-item1.zst (5/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample103-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: 0.007s - ------------------------------------------------------------- -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: 13,172B, BPFP=1.1185 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,284B, BPFP=3.5058 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,856B, BPFP=2.1957 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,932B, BPFP=3.8156 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,364B, BPFP=2.6634 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,396B, BPFP=3.8550 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,020B, BPFP=2.8040 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,148B, BPFP=3.7490 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,388B, BPFP=2.1559 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,028B, BPFP=3.9086 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 110,812B, BPFP=2.3525 -⌛️ [2/4] FRONTEND: Frontend time: 1.796s (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: 1.337s - -[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 0.03121759 4.40198218 - layer.0.v_cache 0.00000027 0.00006261 - layer.1.k_cache 0.00337054 0.33303825 - layer.1.v_cache 0.00000081 0.00019939 - layer.2.k_cache 0.00113856 0.16824770 - layer.2.v_cache 0.00000109 0.00027973 - layer.3.k_cache 0.00130416 0.19704914 - layer.3.v_cache 0.00000204 0.00047568 - layer.4.k_cache 0.00326587 0.41960024 - layer.4.v_cache 0.00000304 0.00076696 - layer.4.output 0.00017424 0.04498262 - ------------------------------------------------------------------------------------- - TOTAL 0.00292864 0.40725945 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 461400 -BPFP 2.7987 bits/point -EBPFP 5.5973 equivalent bits/point -MSE 0.407259 ----------------------- -------------------------------------------------------- -Time: 3.140s Load: 0.007s, Pack+Encode: 1.796s, Decode+Unpack: 1.337s ----------------------- -------------------------------------------------------- -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 0.4073 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample103-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample103-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample105-layer4-item1.zst (6/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample105-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 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, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,944B, BPFP=1.1113 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,124B, BPFP=3.5306 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,732B, BPFP=2.0374 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,088B, BPFP=3.7850 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,840B, BPFP=2.5618 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,296B, BPFP=3.8029 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,524B, BPFP=2.7922 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,876B, BPFP=3.6810 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,992B, BPFP=2.0598 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,292B, BPFP=3.8884 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,476B, BPFP=2.2853 -⌛️ [2/4] FRONTEND: Frontend time: 1.809s (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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.366s - -[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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03135960 4.49631475 - layer.0.v_cache 0.00000027 0.00006369 - layer.1.k_cache 0.00345943 0.32243062 - layer.1.v_cache 0.00000081 0.00019539 - layer.2.k_cache 0.00114090 0.16475740 - layer.2.v_cache 0.00000105 0.00026881 - layer.3.k_cache 0.00133425 0.19644727 - layer.3.v_cache 0.00000201 0.00045841 - layer.4.k_cache 0.00327424 0.41212899 - layer.4.v_cache 0.00000301 0.00072830 - layer.4.output 0.00017460 0.03102826 - ------------------------------------------------------------------------------------- - TOTAL 0.00294814 0.40842190 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 447184 -BPFP 2.7422 bits/point -EBPFP 5.4845 equivalent bits/point -MSE 0.408422 ----------------------- -------------------------------------------------------- -Time: 3.182s Load: 0.007s, Pack+Encode: 1.809s, Decode+Unpack: 1.366s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4084 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample105-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample105-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample107-layer4-item1.zst (7/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample107-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: 0.006s - ------------------------------------------------------------- -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: 12,892B, BPFP=1.1191 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,264B, BPFP=3.5819 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,752B, BPFP=2.0618 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,840B, BPFP=3.8056 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,976B, BPFP=2.6889 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,884B, BPFP=3.8962 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,056B, BPFP=2.7826 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,248B, BPFP=3.7542 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,484B, BPFP=2.0385 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,156B, BPFP=3.9198 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,844B, BPFP=2.2753 -⌛️ [2/4] FRONTEND: Frontend time: 1.771s (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: 1.333s - -[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 0.03118788 4.75419651 - layer.0.v_cache 0.00000027 0.00006464 - layer.1.k_cache 0.00346696 0.33469628 - layer.1.v_cache 0.00000079 0.00019940 - layer.2.k_cache 0.00116280 0.17238602 - layer.2.v_cache 0.00000105 0.00027524 - layer.3.k_cache 0.00130692 0.20037460 - layer.3.v_cache 0.00000207 0.00044722 - layer.4.k_cache 0.00320606 0.39508959 - layer.4.v_cache 0.00000305 0.00075373 - layer.4.output 0.00017695 0.05291052 - ------------------------------------------------------------------------------------- - TOTAL 0.00293183 0.43358038 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 446396 -BPFP 2.7678 bits/point -EBPFP 5.5357 equivalent bits/point -MSE 0.433580 ----------------------- -------------------------------------------------------- -Time: 3.110s Load: 0.006s, Pack+Encode: 1.771s, Decode+Unpack: 1.333s ----------------------- -------------------------------------------------------- -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 0.4336 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample107-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample107-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample111-layer4-item1.zst (8/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample111-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: 0.007s - ------------------------------------------------------------- -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: 13,080B, BPFP=1.1107 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,840B, BPFP=3.5530 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,656B, BPFP=2.1787 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,060B, BPFP=3.8264 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,560B, BPFP=2.6800 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,244B, BPFP=3.8421 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,896B, BPFP=2.7935 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,712B, BPFP=3.7969 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,888B, BPFP=2.1135 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,312B, BPFP=3.9327 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,812B, BPFP=2.3100 -⌛️ [2/4] FRONTEND: Frontend time: 1.808s (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: 1.355s - -[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 0.03125828 4.32054337 - layer.0.v_cache 0.00000028 0.00006308 - layer.1.k_cache 0.00363872 0.35627697 - layer.1.v_cache 0.00000085 0.00019963 - layer.2.k_cache 0.00115386 0.16960145 - layer.2.v_cache 0.00000106 0.00027129 - layer.3.k_cache 0.00130545 0.19742319 - layer.3.v_cache 0.00000205 0.00046490 - layer.4.k_cache 0.00323476 0.40779441 - layer.4.v_cache 0.00000303 0.00073348 - layer.4.output 0.00019770 0.04979793 - ------------------------------------------------------------------------------------- - TOTAL 0.00295637 0.40375453 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 460060 -BPFP 2.7905 bits/point -EBPFP 5.5811 equivalent bits/point -MSE 0.403755 ----------------------- -------------------------------------------------------- -Time: 3.171s Load: 0.007s, Pack+Encode: 1.808s, Decode+Unpack: 1.355s ----------------------- -------------------------------------------------------- -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 0.4038 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample111-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample111-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample112-layer4-item1.zst (9/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample112-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: 12,604B, BPFP=1.1318 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,496B, BPFP=3.6365 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,436B, BPFP=2.1045 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,000B, BPFP=3.7716 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,172B, BPFP=2.6196 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,728B, BPFP=3.9267 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,184B, BPFP=2.7105 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,328B, BPFP=3.8010 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,744B, BPFP=2.0424 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,412B, BPFP=3.8983 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,900B, BPFP=2.3325 -⌛️ [2/4] FRONTEND: Frontend time: 1.749s (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: 1.352s - -[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 0.03173990 4.53349217 - layer.0.v_cache 0.00000027 0.00006360 - layer.1.k_cache 0.00353239 0.34430973 - layer.1.v_cache 0.00000080 0.00020348 - layer.2.k_cache 0.00117039 0.16960177 - layer.2.v_cache 0.00000104 0.00027591 - layer.3.k_cache 0.00134488 0.19177527 - layer.3.v_cache 0.00000203 0.00047129 - layer.4.k_cache 0.00324251 0.38955724 - layer.4.v_cache 0.00000295 0.00076472 - layer.4.output 0.00019562 0.04541811 - ------------------------------------------------------------------------------------- - TOTAL 0.00298712 0.41515626 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 434004 -BPFP 2.7838 bits/point -EBPFP 5.5676 equivalent bits/point -MSE 0.415156 ----------------------- -------------------------------------------------------- -Time: 3.107s Load: 0.006s, Pack+Encode: 1.749s, Decode+Unpack: 1.352s ----------------------- -------------------------------------------------------- -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 0.4152 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample112-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample112-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample113-layer4-item1.zst (10/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample113-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: 12,664B, BPFP=1.1372 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,280B, BPFP=3.6171 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,552B, BPFP=2.1149 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,124B, BPFP=3.7827 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,540B, BPFP=2.5629 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,384B, BPFP=3.8958 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,368B, BPFP=2.7270 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,180B, BPFP=3.7877 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,700B, BPFP=2.0384 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,428B, BPFP=3.8998 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 102,620B, BPFP=2.3038 -⌛️ [2/4] FRONTEND: Frontend time: 1.760s (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: 1.384s - -[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 0.03084559 4.42993304 - layer.0.v_cache 0.00000028 0.00006446 - layer.1.k_cache 0.00352810 0.32485848 - layer.1.v_cache 0.00000080 0.00020588 - layer.2.k_cache 0.00113435 0.16701509 - layer.2.v_cache 0.00000106 0.00028442 - layer.3.k_cache 0.00133457 0.19276240 - layer.3.v_cache 0.00000203 0.00047595 - layer.4.k_cache 0.00324423 0.38845457 - layer.4.v_cache 0.00000296 0.00073961 - layer.4.output 0.00017424 0.04213268 - ------------------------------------------------------------------------------------- - TOTAL 0.00291364 0.40523747 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 431840 -BPFP 2.7699 bits/point -EBPFP 5.5398 equivalent bits/point -MSE 0.405237 ----------------------- -------------------------------------------------------- -Time: 3.151s Load: 0.006s, Pack+Encode: 1.760s, Decode+Unpack: 1.384s ----------------------- -------------------------------------------------------- -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 0.4052 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample113-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample113-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample117-layer4-item1.zst (11/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample117-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: 12,484B, BPFP=1.1474 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 39,736B, BPFP=3.6522 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,968B, BPFP=2.1110 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,112B, BPFP=3.7787 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,064B, BPFP=2.5794 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,668B, BPFP=3.9217 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,820B, BPFP=2.6489 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,948B, BPFP=3.7636 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,240B, BPFP=2.0441 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,180B, BPFP=3.8768 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,496B, BPFP=2.3781 -⌛️ [2/4] FRONTEND: Frontend time: 1.772s (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: 1.352s - -[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 0.03172885 4.34730296 - layer.0.v_cache 0.00000027 0.00006448 - layer.1.k_cache 0.00352961 0.33856812 - layer.1.v_cache 0.00000080 0.00020792 - layer.2.k_cache 0.00112280 0.16664084 - layer.2.v_cache 0.00000110 0.00028162 - layer.3.k_cache 0.00133082 0.18712381 - layer.3.v_cache 0.00000207 0.00047166 - layer.4.k_cache 0.00329730 0.40627311 - layer.4.v_cache 0.00000296 0.00076074 - layer.4.output 0.00017366 0.04857035 - ------------------------------------------------------------------------------------- - TOTAL 0.00297937 0.40299833 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 424716 -BPFP 2.7883 bits/point -EBPFP 5.5766 equivalent bits/point -MSE 0.402998 ----------------------- -------------------------------------------------------- -Time: 3.131s Load: 0.006s, Pack+Encode: 1.772s, Decode+Unpack: 1.352s ----------------------- -------------------------------------------------------- -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 0.4030 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample117-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample117-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample12-layer4-item1.zst (12/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample12-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: 0.007s - ------------------------------------------------------------- -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: 12,848B, BPFP=1.0793 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,696B, BPFP=3.5027 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,176B, BPFP=2.1989 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,364B, BPFP=3.8108 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,512B, BPFP=2.6472 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,024B, BPFP=3.7823 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,412B, BPFP=2.7228 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,032B, BPFP=3.6989 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,024B, BPFP=2.0181 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,768B, BPFP=3.8448 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,624B, BPFP=2.2392 -⌛️ [2/4] FRONTEND: Frontend time: 1.786s (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: 1.332s - -[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 0.03136649 4.33496750 - layer.0.v_cache 0.00000028 0.00006268 - layer.1.k_cache 0.00359034 0.36143498 - layer.1.v_cache 0.00000079 0.00019732 - layer.2.k_cache 0.00112294 0.17016786 - layer.2.v_cache 0.00000105 0.00026525 - layer.3.k_cache 0.00134461 0.19680626 - layer.3.v_cache 0.00000211 0.00045355 - layer.4.k_cache 0.00331549 0.39842044 - layer.4.v_cache 0.00000311 0.00076844 - layer.4.output 0.00017890 0.05294790 - ------------------------------------------------------------------------------------- - TOTAL 0.00296163 0.40538113 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 455480 -BPFP 2.7331 bits/point -EBPFP 5.4661 equivalent bits/point -MSE 0.405381 ----------------------- -------------------------------------------------------- -Time: 3.125s Load: 0.007s, Pack+Encode: 1.786s, Decode+Unpack: 1.332s ----------------------- -------------------------------------------------------- -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 0.4054 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample12-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample12-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample121-layer4-item1.zst (13/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample121-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: 12,316B, BPFP=1.1455 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 39,952B, BPFP=3.7158 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,420B, BPFP=2.0852 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,316B, BPFP=3.8426 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,648B, BPFP=2.4784 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,852B, BPFP=3.9855 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,068B, BPFP=2.6105 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,032B, BPFP=3.8162 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,924B, BPFP=2.0391 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,720B, BPFP=3.9732 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 100,668B, BPFP=2.3407 -⌛️ [2/4] FRONTEND: Frontend time: 1.799s (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: 1.358s - -[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 0.03306588 4.27285839 - layer.0.v_cache 0.00000027 0.00006183 - layer.1.k_cache 0.00337037 0.33343744 - layer.1.v_cache 0.00000088 0.00019658 - layer.2.k_cache 0.00113180 0.16495850 - layer.2.v_cache 0.00000103 0.00026718 - layer.3.k_cache 0.00132828 0.18698102 - layer.3.v_cache 0.00000204 0.00045459 - layer.4.k_cache 0.00320824 0.39887901 - layer.4.v_cache 0.00000300 0.00076167 - layer.4.output 0.00017823 0.03779923 - ------------------------------------------------------------------------------------- - TOTAL 0.00305891 0.39357522 - (elements=1,204,224) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1204224 -Total Bytes 419916 -BPFP 2.7896 bits/point -EBPFP 5.5792 equivalent bits/point -MSE 0.393575 ----------------------- -------------------------------------------------------- -Time: 3.163s Load: 0.006s, Pack+Encode: 1.799s, Decode+Unpack: 1.358s ----------------------- -------------------------------------------------------- -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 0.3936 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample121-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample121-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample123-layer4-item1.zst (14/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample123-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: 0.007s - ------------------------------------------------------------- -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: 12,860B, BPFP=1.1163 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,008B, BPFP=3.5597 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,588B, BPFP=2.0476 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,840B, BPFP=3.8056 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,672B, BPFP=2.6625 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,540B, BPFP=3.8663 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,612B, BPFP=2.7441 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,336B, BPFP=3.7618 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,556B, BPFP=2.0448 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,780B, BPFP=3.8872 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,716B, BPFP=2.2942 -⌛️ [2/4] FRONTEND: Frontend time: 1.783s (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: 1.367s - -[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 0.03166741 4.56743300 - layer.0.v_cache 0.00000027 0.00006416 - layer.1.k_cache 0.00338869 0.32903540 - layer.1.v_cache 0.00000080 0.00019955 - layer.2.k_cache 0.00112306 0.16942819 - layer.2.v_cache 0.00000109 0.00027365 - layer.3.k_cache 0.00129226 0.19648760 - layer.3.v_cache 0.00000208 0.00045194 - layer.4.k_cache 0.00325011 0.39091237 - layer.4.v_cache 0.00000309 0.00075695 - layer.4.output 0.00016397 0.04131566 - ------------------------------------------------------------------------------------- - TOTAL 0.00295605 0.41573610 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 445508 -BPFP 2.7623 bits/point -EBPFP 5.5247 equivalent bits/point -MSE 0.415736 ----------------------- -------------------------------------------------------- -Time: 3.158s Load: 0.007s, Pack+Encode: 1.783s, Decode+Unpack: 1.367s ----------------------- -------------------------------------------------------- -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 0.4157 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample123-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample123-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample124-layer4-item1.zst (15/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample124-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 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, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,688B, BPFP=1.1264 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,500B, BPFP=3.5955 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,300B, BPFP=2.0685 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,888B, BPFP=3.8075 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,004B, BPFP=2.5749 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,408B, BPFP=3.8537 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,812B, BPFP=2.7354 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,936B, BPFP=3.7230 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,900B, BPFP=2.0330 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,428B, BPFP=3.8555 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,776B, BPFP=2.3477 -⌛️ [2/4] FRONTEND: Frontend time: 1.817s (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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.355s - -[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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03171545 4.53990520 - layer.0.v_cache 0.00000027 0.00006461 - layer.1.k_cache 0.00364625 0.33600354 - layer.1.v_cache 0.00000082 0.00020473 - layer.2.k_cache 0.00114710 0.16767043 - layer.2.v_cache 0.00000106 0.00027362 - layer.3.k_cache 0.00132811 0.20007805 - layer.3.v_cache 0.00000203 0.00046852 - layer.4.k_cache 0.00332194 0.39926035 - layer.4.v_cache 0.00000300 0.00073844 - layer.4.output 0.00016865 0.04019246 - ------------------------------------------------------------------------------------- - TOTAL 0.00298862 0.41467410 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 436640 -BPFP 2.7689 bits/point -EBPFP 5.5377 equivalent bits/point -MSE 0.414674 ----------------------- -------------------------------------------------------- -Time: 3.178s Load: 0.006s, Pack+Encode: 1.817s, Decode+Unpack: 1.355s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4147 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample124-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample124-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample126-layer4-item1.zst (16/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample126-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: 12,900B, BPFP=1.1324 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,696B, BPFP=3.5723 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,824B, BPFP=2.0913 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,416B, BPFP=3.8111 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,208B, BPFP=2.5639 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,876B, BPFP=3.8515 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,252B, BPFP=2.7433 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,208B, BPFP=3.7051 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,228B, BPFP=2.0390 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,864B, BPFP=3.8504 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,876B, BPFP=2.3235 -⌛️ [2/4] FRONTEND: Frontend time: 1.788s (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: 1.354s - -[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 0.03218093 4.68383172 - layer.0.v_cache 0.00000027 0.00006135 - layer.1.k_cache 0.00364370 0.35047503 - layer.1.v_cache 0.00000081 0.00021348 - layer.2.k_cache 0.00115129 0.17166017 - layer.2.v_cache 0.00000108 0.00028985 - layer.3.k_cache 0.00133425 0.19800842 - layer.3.v_cache 0.00000206 0.00048953 - layer.4.k_cache 0.00326870 0.40676121 - layer.4.v_cache 0.00000299 0.00077185 - layer.4.output 0.00018569 0.04860782 - ------------------------------------------------------------------------------------- - TOTAL 0.00302349 0.42907099 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 440348 -BPFP 2.7610 bits/point -EBPFP 5.5220 equivalent bits/point -MSE 0.429071 ----------------------- -------------------------------------------------------- -Time: 3.148s Load: 0.007s, Pack+Encode: 1.788s, Decode+Unpack: 1.354s ----------------------- -------------------------------------------------------- -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 0.4291 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample126-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample126-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample13-layer4-item1.zst (17/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample13-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: 0.007s - ------------------------------------------------------------- -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: 12,880B, BPFP=1.0820 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,444B, BPFP=3.4815 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,936B, BPFP=2.1788 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,112B, BPFP=3.7897 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,504B, BPFP=2.6465 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,912B, BPFP=3.7728 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,360B, BPFP=2.7184 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,304B, BPFP=3.7218 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,408B, BPFP=2.0504 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,600B, BPFP=3.8306 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,996B, BPFP=2.2261 -⌛️ [2/4] FRONTEND: Frontend time: 1.757s (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: 1.357s - -[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 0.03082583 4.17382353 - layer.0.v_cache 0.00000028 0.00006130 - layer.1.k_cache 0.00340863 0.36907532 - layer.1.v_cache 0.00000080 0.00019379 - layer.2.k_cache 0.00115869 0.16994156 - layer.2.v_cache 0.00000105 0.00026600 - layer.3.k_cache 0.00132054 0.19066550 - layer.3.v_cache 0.00000208 0.00045193 - layer.4.k_cache 0.00325612 0.39799984 - layer.4.v_cache 0.00000315 0.00077097 - layer.4.output 0.00020065 0.05895423 - ------------------------------------------------------------------------------------- - TOTAL 0.00291284 0.39564762 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 454456 -BPFP 2.7269 bits/point -EBPFP 5.4538 equivalent bits/point -MSE 0.395648 ----------------------- -------------------------------------------------------- -Time: 3.120s Load: 0.007s, Pack+Encode: 1.757s, Decode+Unpack: 1.357s ----------------------- -------------------------------------------------------- -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 0.3956 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample13-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample13-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample132-layer4-item1.zst (18/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample132-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: 0.008s - ------------------------------------------------------------- -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: 12,644B, BPFP=1.0976 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,036B, BPFP=3.5622 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,180B, BPFP=2.0990 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,932B, BPFP=3.8135 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,988B, BPFP=2.6031 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,956B, BPFP=3.8156 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,028B, BPFP=2.7802 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,404B, BPFP=3.7677 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,772B, BPFP=2.0635 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,832B, BPFP=3.8917 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,892B, BPFP=2.2546 -⌛️ [2/4] FRONTEND: Frontend time: 1.777s (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: 1.340s - -[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 0.03096869 4.45781657 - layer.0.v_cache 0.00000027 0.00006361 - layer.1.k_cache 0.00338461 0.35461593 - layer.1.v_cache 0.00000079 0.00019854 - layer.2.k_cache 0.00111556 0.16928139 - layer.2.v_cache 0.00000105 0.00027069 - layer.3.k_cache 0.00129967 0.19924151 - layer.3.v_cache 0.00000204 0.00044604 - layer.4.k_cache 0.00324687 0.41496836 - layer.4.v_cache 0.00000301 0.00072842 - layer.4.output 0.00020513 0.04694221 - ------------------------------------------------------------------------------------- - TOTAL 0.00291736 0.41324285 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 443664 -BPFP 2.7509 bits/point -EBPFP 5.5018 equivalent bits/point -MSE 0.413243 ----------------------- -------------------------------------------------------- -Time: 3.125s Load: 0.008s, Pack+Encode: 1.777s, Decode+Unpack: 1.340s ----------------------- -------------------------------------------------------- -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 0.4132 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample132-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample132-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample134-layer4-item1.zst (19/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample134-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: 12,552B, BPFP=1.1018 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,976B, BPFP=3.5969 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,068B, BPFP=2.1127 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,600B, BPFP=3.8272 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,996B, BPFP=2.5453 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,076B, BPFP=3.8690 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,276B, BPFP=2.6577 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,196B, BPFP=3.7040 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,372B, BPFP=2.0516 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,044B, BPFP=3.8662 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,960B, BPFP=2.3034 -⌛️ [2/4] FRONTEND: Frontend time: 1.819s (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: 1.350s - -[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 0.03072730 4.60989620 - layer.0.v_cache 0.00000028 0.00006199 - layer.1.k_cache 0.00346408 0.33999276 - layer.1.v_cache 0.00000080 0.00020328 - layer.2.k_cache 0.00114708 0.16869183 - layer.2.v_cache 0.00000106 0.00028125 - layer.3.k_cache 0.00132643 0.19006892 - layer.3.v_cache 0.00000205 0.00047705 - layer.4.k_cache 0.00323342 0.39954509 - layer.4.v_cache 0.00000308 0.00078104 - layer.4.output 0.00016383 0.04127937 - ------------------------------------------------------------------------------------- - TOTAL 0.00289721 0.41965121 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 439116 -BPFP 2.7533 bits/point -EBPFP 5.5066 equivalent bits/point -MSE 0.419651 ----------------------- -------------------------------------------------------- -Time: 3.175s Load: 0.006s, Pack+Encode: 1.819s, Decode+Unpack: 1.350s ----------------------- -------------------------------------------------------- -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 0.4197 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample134-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample134-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample135-layer4-item1.zst (20/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample135-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: 0.007s - ------------------------------------------------------------- -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: 13,032B, BPFP=1.1067 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,460B, BPFP=3.5207 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,004B, BPFP=2.1233 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,524B, BPFP=3.7809 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,596B, BPFP=2.5982 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,144B, BPFP=3.8336 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,600B, BPFP=2.7683 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,836B, BPFP=3.7225 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,572B, BPFP=2.0866 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,560B, BPFP=3.9538 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,012B, BPFP=2.2506 -⌛️ [2/4] FRONTEND: Frontend time: 1.798s (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: 1.313s - -[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 0.03070580 4.48114080 - layer.0.v_cache 0.00000027 0.00006431 - layer.1.k_cache 0.00347185 0.34942196 - layer.1.v_cache 0.00000080 0.00019774 - layer.2.k_cache 0.00112340 0.16934106 - layer.2.v_cache 0.00000106 0.00027146 - layer.3.k_cache 0.00133882 0.19828299 - layer.3.v_cache 0.00000200 0.00045583 - layer.4.k_cache 0.00333409 0.41202330 - layer.4.v_cache 0.00000307 0.00075124 - layer.4.output 0.00016236 0.05306143 - ------------------------------------------------------------------------------------- - TOTAL 0.00290219 0.41601403 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 453340 -BPFP 2.7498 bits/point -EBPFP 5.4996 equivalent bits/point -MSE 0.416014 ----------------------- -------------------------------------------------------- -Time: 3.118s Load: 0.007s, Pack+Encode: 1.798s, Decode+Unpack: 1.313s ----------------------- -------------------------------------------------------- -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 0.4160 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample135-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample135-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample136-layer4-item1.zst (21/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample136-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: 0.007s - ------------------------------------------------------------- -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: 12,684B, BPFP=1.1390 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,284B, BPFP=3.6175 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,224B, BPFP=2.0855 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,004B, BPFP=3.8617 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,288B, BPFP=2.5402 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,624B, BPFP=3.9174 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,668B, BPFP=2.7540 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,072B, BPFP=3.7780 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,600B, BPFP=2.0295 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,580B, BPFP=3.9134 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 102,436B, BPFP=2.2997 -⌛️ [2/4] FRONTEND: Frontend time: 1.803s (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: 1.346s - -[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 0.03132415 4.46646013 - layer.0.v_cache 0.00000028 0.00006203 - layer.1.k_cache 0.00348164 0.32427330 - layer.1.v_cache 0.00000079 0.00020143 - layer.2.k_cache 0.00113511 0.16296220 - layer.2.v_cache 0.00000106 0.00027370 - layer.3.k_cache 0.00131762 0.19099955 - layer.3.v_cache 0.00000207 0.00045875 - layer.4.k_cache 0.00320374 0.38220031 - layer.4.v_cache 0.00000303 0.00075571 - layer.4.output 0.00016514 0.03824967 - ------------------------------------------------------------------------------------- - TOTAL 0.00293786 0.40583184 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 432464 -BPFP 2.7739 bits/point -EBPFP 5.5478 equivalent bits/point -MSE 0.405832 ----------------------- -------------------------------------------------------- -Time: 3.156s Load: 0.007s, Pack+Encode: 1.803s, Decode+Unpack: 1.346s ----------------------- -------------------------------------------------------- -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 0.4058 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample136-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample136-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample137-layer4-item1.zst (22/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample137-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 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, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,676B, BPFP=1.1254 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,064B, BPFP=3.5568 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,404B, BPFP=2.0778 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,340B, BPFP=3.8477 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,384B, BPFP=2.5199 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,896B, BPFP=3.8082 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,800B, BPFP=2.7344 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,312B, BPFP=3.6676 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,120B, BPFP=2.0526 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,504B, BPFP=3.7734 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,836B, BPFP=2.3046 -⌛️ [2/4] FRONTEND: Frontend time: 1.798s (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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.214s - -[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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03137242 4.36708658 - layer.0.v_cache 0.00000027 0.00006216 - layer.1.k_cache 0.00337550 0.33520675 - layer.1.v_cache 0.00000078 0.00020165 - layer.2.k_cache 0.00112765 0.16873169 - layer.2.v_cache 0.00000106 0.00027424 - layer.3.k_cache 0.00130259 0.19736947 - layer.3.v_cache 0.00000201 0.00046033 - layer.4.k_cache 0.00311145 0.39987902 - layer.4.v_cache 0.00000304 0.00077110 - layer.4.output 0.00020755 0.04422792 - ------------------------------------------------------------------------------------- - TOTAL 0.00293764 0.40335390 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 432336 -BPFP 2.7416 bits/point -EBPFP 5.4832 equivalent bits/point -MSE 0.403354 ----------------------- -------------------------------------------------------- -Time: 3.019s Load: 0.006s, Pack+Encode: 1.798s, Decode+Unpack: 1.214s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4034 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample137-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample137-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample138-layer4-item1.zst (23/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample138-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: 12,756B, BPFP=1.0602 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,124B, BPFP=3.4179 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,788B, BPFP=2.1433 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,980B, BPFP=3.7384 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,172B, BPFP=2.6739 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,104B, BPFP=3.7487 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,624B, BPFP=2.7114 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,004B, BPFP=3.6572 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,448B, BPFP=2.1150 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,716B, BPFP=3.7995 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,664B, BPFP=2.1955 -⌛️ [2/4] FRONTEND: Frontend time: 1.706s (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: 1.334s - -[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 0.03043337 4.43442617 - layer.0.v_cache 0.00000027 0.00006349 - layer.1.k_cache 0.00337587 0.34733293 - layer.1.v_cache 0.00000080 0.00020299 - layer.2.k_cache 0.00115990 0.17095888 - layer.2.v_cache 0.00000105 0.00027223 - layer.3.k_cache 0.00132055 0.19476148 - layer.3.v_cache 0.00000202 0.00045671 - layer.4.k_cache 0.00326599 0.41439941 - layer.4.v_cache 0.00000301 0.00074612 - layer.4.output 0.00017344 0.05147430 - ------------------------------------------------------------------------------------- - TOTAL 0.00287547 0.41210840 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 455380 -BPFP 2.7034 bits/point -EBPFP 5.4068 equivalent bits/point -MSE 0.412108 ----------------------- -------------------------------------------------------- -Time: 3.046s Load: 0.006s, Pack+Encode: 1.706s, Decode+Unpack: 1.334s ----------------------- -------------------------------------------------------- -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 0.4121 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample138-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample138-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample139-layer4-item1.zst (24/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample139-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 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, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,720B, BPFP=1.1293 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,580B, BPFP=3.6026 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,624B, BPFP=2.0973 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,344B, BPFP=3.8480 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,844B, BPFP=2.5607 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,404B, BPFP=3.8533 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,928B, BPFP=2.7457 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,608B, BPFP=3.7827 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,860B, BPFP=2.0295 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,484B, BPFP=3.8604 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,048B, BPFP=2.3093 -⌛️ [2/4] FRONTEND: Frontend time: 1.821s (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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.349s - -[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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03171148 4.35918427 - layer.0.v_cache 0.00000028 0.00006367 - layer.1.k_cache 0.00359844 0.33813607 - layer.1.v_cache 0.00000084 0.00020244 - layer.2.k_cache 0.00114211 0.16716053 - layer.2.v_cache 0.00000108 0.00028953 - layer.3.k_cache 0.00132976 0.20397102 - layer.3.v_cache 0.00000213 0.00048540 - layer.4.k_cache 0.00324865 0.39968881 - layer.4.v_cache 0.00000293 0.00073980 - layer.4.output 0.00016674 0.03571143 - ------------------------------------------------------------------------------------- - TOTAL 0.00297890 0.40091195 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 436444 -BPFP 2.7676 bits/point -EBPFP 5.5353 equivalent bits/point -MSE 0.400912 ----------------------- -------------------------------------------------------- -Time: 3.176s Load: 0.006s, Pack+Encode: 1.821s, Decode+Unpack: 1.349s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4009 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample139-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample139-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample140-layer4-item1.zst (25/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample140-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: 12,712B, BPFP=1.1159 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,552B, BPFP=3.6475 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,540B, BPFP=2.0664 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,992B, BPFP=3.8617 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,492B, BPFP=2.5888 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,016B, BPFP=3.8638 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,476B, BPFP=2.6752 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,996B, BPFP=3.6864 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,220B, BPFP=2.0383 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,640B, BPFP=3.8308 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,784B, BPFP=2.3434 -⌛️ [2/4] FRONTEND: Frontend time: 1.841s (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: 1.363s - -[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 0.03094440 4.33491636 - layer.0.v_cache 0.00000027 0.00006129 - layer.1.k_cache 0.00351605 0.33741293 - layer.1.v_cache 0.00000081 0.00019818 - layer.2.k_cache 0.00112935 0.17144600 - layer.2.v_cache 0.00000105 0.00026655 - layer.3.k_cache 0.00132761 0.19592671 - layer.3.v_cache 0.00000216 0.00047017 - layer.4.k_cache 0.00319280 0.39468787 - layer.4.v_cache 0.00000300 0.00076475 - layer.4.output 0.00020683 0.04737241 - ------------------------------------------------------------------------------------- - TOTAL 0.00292463 0.40183146 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 441420 -BPFP 2.7677 bits/point -EBPFP 5.5355 equivalent bits/point -MSE 0.401831 ----------------------- -------------------------------------------------------- -Time: 3.209s Load: 0.006s, Pack+Encode: 1.841s, Decode+Unpack: 1.363s ----------------------- -------------------------------------------------------- -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 0.4018 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample140-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample140-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample141-layer4-item1.zst (26/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample141-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: 0.008s - ------------------------------------------------------------- -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: 13,052B, BPFP=1.1084 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,940B, BPFP=3.5615 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,528B, BPFP=2.1678 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,652B, BPFP=3.7918 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,056B, BPFP=2.6372 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,400B, BPFP=3.8553 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,632B, BPFP=2.7711 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,328B, BPFP=3.7643 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,232B, BPFP=2.0577 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,820B, BPFP=3.8910 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,792B, BPFP=2.2884 -⌛️ [2/4] FRONTEND: Frontend time: 1.797s (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: 1.363s - -[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 0.03142291 4.39088042 - layer.0.v_cache 0.00000027 0.00006410 - layer.1.k_cache 0.00332372 0.33385119 - layer.1.v_cache 0.00000083 0.00019250 - layer.2.k_cache 0.00117076 0.17047289 - layer.2.v_cache 0.00000109 0.00028047 - layer.3.k_cache 0.00132340 0.19714766 - layer.3.v_cache 0.00000203 0.00044254 - layer.4.k_cache 0.00333891 0.40275682 - layer.4.v_cache 0.00000309 0.00076049 - layer.4.output 0.00017306 0.03469711 - ------------------------------------------------------------------------------------- - TOTAL 0.00294852 0.40254554 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 456432 -BPFP 2.7685 bits/point -EBPFP 5.5371 equivalent bits/point -MSE 0.402546 ----------------------- -------------------------------------------------------- -Time: 3.168s Load: 0.008s, Pack+Encode: 1.797s, Decode+Unpack: 1.363s ----------------------- -------------------------------------------------------- -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 0.4025 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample141-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample141-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample142-layer4-item1.zst (27/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample142-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: 12,956B, BPFP=1.0884 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,296B, BPFP=3.4691 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,532B, BPFP=2.1448 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,144B, BPFP=3.7923 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,568B, BPFP=2.6519 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,208B, BPFP=3.7977 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,424B, BPFP=2.7238 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,704B, BPFP=3.7554 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,388B, BPFP=2.0487 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,804B, BPFP=3.8478 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,516B, BPFP=2.2790 -⌛️ [2/4] FRONTEND: Frontend time: 1.789s (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: 1.385s - -[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, 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 0.03274377 4.28416853 - layer.0.v_cache 0.00000028 0.00006523 - layer.1.k_cache 0.00350561 0.35921208 - layer.1.v_cache 0.00000082 0.00020655 - layer.2.k_cache 0.00114290 0.17210930 - layer.2.v_cache 0.00000109 0.00028109 - layer.3.k_cache 0.00132787 0.19621244 - layer.3.v_cache 0.00000209 0.00048894 - layer.4.k_cache 0.00329855 0.39898813 - layer.4.v_cache 0.00000319 0.00077284 - layer.4.output 0.00017541 0.05325955 - ------------------------------------------------------------------------------------- - TOTAL 0.00305199 0.40182452 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 457540 -BPFP 2.7454 bits/point -EBPFP 5.4908 equivalent bits/point -MSE 0.401825 ----------------------- -------------------------------------------------------- -Time: 3.180s Load: 0.006s, Pack+Encode: 1.789s, Decode+Unpack: 1.385s ----------------------- -------------------------------------------------------- -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 0.4018 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample142-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample142-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample145-layer4-item1.zst (28/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample145-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: 0.005s - ------------------------------------------------------------- -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: 12,072B, BPFP=1.1502 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 38,512B, BPFP=3.6692 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,060B, BPFP=2.1018 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,540B, BPFP=3.9577 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,508B, BPFP=2.5255 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 41,484B, BPFP=3.9524 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,928B, BPFP=2.6608 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,172B, BPFP=3.8274 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,664B, BPFP=2.0640 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 40,944B, BPFP=3.9009 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 101,604B, BPFP=2.4201 -⌛️ [2/4] FRONTEND: Frontend time: 1.804s (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: 1.341s - -[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 0.03240362 4.60537385 - layer.0.v_cache 0.00000028 0.00006483 - layer.1.k_cache 0.00358562 0.34682269 - layer.1.v_cache 0.00000082 0.00020661 - layer.2.k_cache 0.00116773 0.16286025 - layer.2.v_cache 0.00000108 0.00028731 - layer.3.k_cache 0.00133166 0.19243282 - layer.3.v_cache 0.00000211 0.00047012 - layer.4.k_cache 0.00321141 0.39420495 - layer.4.v_cache 0.00000303 0.00076534 - layer.4.output 0.00019620 0.04561997 - ------------------------------------------------------------------------------------- - TOTAL 0.00303515 0.42042633 - (elements=1,175,552) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1175552 -Total Bytes 414488 -BPFP 2.8207 bits/point -EBPFP 5.6414 equivalent bits/point -MSE 0.420426 ----------------------- -------------------------------------------------------- -Time: 3.149s Load: 0.005s, Pack+Encode: 1.804s, Decode+Unpack: 1.341s ----------------------- -------------------------------------------------------- -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 0.4204 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample145-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample145-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample147-layer4-item1.zst (29/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample147-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: 12,836B, BPFP=1.1527 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,404B, BPFP=3.6282 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,396B, BPFP=2.1009 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,376B, BPFP=3.8053 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,008B, BPFP=2.6049 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,132B, BPFP=3.8732 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,776B, BPFP=2.7636 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,688B, BPFP=3.7435 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,900B, BPFP=2.0564 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,236B, BPFP=3.8825 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,052B, BPFP=2.3808 -⌛️ [2/4] FRONTEND: Frontend time: 1.755s (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: 1.346s - -[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 0.03155136 4.44211692 - layer.0.v_cache 0.00000028 0.00006423 - layer.1.k_cache 0.00336733 0.31975117 - layer.1.v_cache 0.00000080 0.00020756 - layer.2.k_cache 0.00113514 0.16226738 - layer.2.v_cache 0.00000108 0.00027918 - layer.3.k_cache 0.00133108 0.19431809 - layer.3.v_cache 0.00000214 0.00047181 - layer.4.k_cache 0.00321453 0.38636245 - layer.4.v_cache 0.00000314 0.00074197 - layer.4.output 0.00017293 0.04644746 - ------------------------------------------------------------------------------------- - TOTAL 0.00294990 0.40659790 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 435804 -BPFP 2.7953 bits/point -EBPFP 5.5907 equivalent bits/point -MSE 0.406598 ----------------------- -------------------------------------------------------- -Time: 3.107s Load: 0.006s, Pack+Encode: 1.755s, Decode+Unpack: 1.346s ----------------------- -------------------------------------------------------- -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 0.4066 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample147-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample147-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample149-layer4-item1.zst (30/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample149-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: 0.007s - ------------------------------------------------------------- -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: 12,968B, BPFP=1.1012 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,276B, BPFP=3.5051 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,892B, BPFP=2.1138 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,324B, BPFP=3.8488 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,644B, BPFP=2.6022 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,288B, BPFP=3.8458 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,336B, BPFP=2.7459 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,940B, BPFP=3.7313 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,496B, BPFP=2.0802 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,232B, BPFP=3.9260 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,988B, BPFP=2.2501 -⌛️ [2/4] FRONTEND: Frontend time: 1.756s (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: 1.340s - -[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 0.03082907 4.61912669 - layer.0.v_cache 0.00000028 0.00006399 - layer.1.k_cache 0.00336151 0.34830317 - layer.1.v_cache 0.00000078 0.00019250 - layer.2.k_cache 0.00113870 0.16537438 - layer.2.v_cache 0.00000105 0.00026972 - layer.3.k_cache 0.00130133 0.19470646 - layer.3.v_cache 0.00000208 0.00045913 - layer.4.k_cache 0.00328680 0.40724588 - layer.4.v_cache 0.00000308 0.00075282 - layer.4.output 0.00016380 0.03550826 - ------------------------------------------------------------------------------------- - TOTAL 0.00289856 0.41989484 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 453384 -BPFP 2.7500 bits/point -EBPFP 5.5001 equivalent bits/point -MSE 0.419895 ----------------------- -------------------------------------------------------- -Time: 3.102s Load: 0.007s, Pack+Encode: 1.756s, Decode+Unpack: 1.340s ----------------------- -------------------------------------------------------- -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 0.4199 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample149-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample149-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample15-layer4-item1.zst (31/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample15-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: 0.007s - ------------------------------------------------------------- -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: 12,908B, BPFP=1.1205 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,056B, BPFP=3.5639 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,360B, BPFP=2.1146 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,048B, BPFP=3.8236 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,252B, BPFP=2.6260 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,444B, BPFP=3.8580 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,084B, BPFP=2.7851 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,964B, BPFP=3.7295 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,048B, BPFP=2.0007 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,860B, BPFP=3.8941 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,588B, BPFP=2.2914 -⌛️ [2/4] FRONTEND: Frontend time: 1.782s (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: 1.367s - -[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 0.03195141 4.34965346 - layer.0.v_cache 0.00000027 0.00006255 - layer.1.k_cache 0.00349494 0.34374457 - layer.1.v_cache 0.00000082 0.00019843 - layer.2.k_cache 0.00113744 0.17284134 - layer.2.v_cache 0.00000112 0.00027840 - layer.3.k_cache 0.00129971 0.20012184 - layer.3.v_cache 0.00000221 0.00047136 - layer.4.k_cache 0.00331958 0.39740533 - layer.4.v_cache 0.00000310 0.00076173 - layer.4.output 0.00017227 0.04771345 - ------------------------------------------------------------------------------------- - TOTAL 0.00299284 0.40402806 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 445612 -BPFP 2.7630 bits/point -EBPFP 5.5259 equivalent bits/point -MSE 0.404028 ----------------------- -------------------------------------------------------- -Time: 3.156s Load: 0.007s, Pack+Encode: 1.782s, Decode+Unpack: 1.367s ----------------------- -------------------------------------------------------- -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 0.4040 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample15-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample15-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample153-layer4-item1.zst (32/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample153-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: 12,744B, BPFP=1.0592 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,296B, BPFP=3.4322 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,888B, BPFP=2.1516 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,140B, BPFP=3.7517 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,956B, BPFP=2.6559 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,136B, BPFP=3.7513 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,708B, BPFP=2.7184 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,644B, BPFP=3.6273 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,132B, BPFP=2.0888 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,484B, BPFP=3.7803 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,712B, BPFP=2.1965 -⌛️ [2/4] FRONTEND: Frontend time: 1.827s (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: 1.360s - -[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 0.03049540 4.25669796 - layer.0.v_cache 0.00000028 0.00006292 - layer.1.k_cache 0.00333180 0.33636791 - layer.1.v_cache 0.00000080 0.00020227 - layer.2.k_cache 0.00116407 0.17065170 - layer.2.v_cache 0.00000105 0.00027201 - layer.3.k_cache 0.00132929 0.19478023 - layer.3.v_cache 0.00000200 0.00045806 - layer.4.k_cache 0.00323017 0.41753789 - layer.4.v_cache 0.00000303 0.00074694 - layer.4.output 0.00017160 0.05204588 - ------------------------------------------------------------------------------------- - TOTAL 0.00287459 0.39899724 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 454840 -BPFP 2.7002 bits/point -EBPFP 5.4004 equivalent bits/point -MSE 0.398997 ----------------------- -------------------------------------------------------- -Time: 3.193s Load: 0.006s, Pack+Encode: 1.827s, Decode+Unpack: 1.360s ----------------------- -------------------------------------------------------- -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 0.3990 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample153-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample153-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample154-layer4-item1.zst (33/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample154-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: 12,508B, BPFP=1.1496 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 39,452B, BPFP=3.6261 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,992B, BPFP=2.1132 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,732B, BPFP=3.8357 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,620B, BPFP=2.5386 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,476B, BPFP=3.9040 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,700B, BPFP=2.6379 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,980B, BPFP=3.7665 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,208B, BPFP=2.0412 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,652B, BPFP=3.9202 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 101,684B, BPFP=2.3365 -⌛️ [2/4] FRONTEND: Frontend time: 1.751s (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: 1.350s - -[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 0.03223823 4.69268799 - layer.0.v_cache 0.00000028 0.00006373 - layer.1.k_cache 0.00365975 0.32979920 - layer.1.v_cache 0.00000080 0.00020474 - layer.2.k_cache 0.00113736 0.16542320 - layer.2.v_cache 0.00000107 0.00027989 - layer.3.k_cache 0.00130376 0.18657372 - layer.3.v_cache 0.00000210 0.00046492 - layer.4.k_cache 0.00330933 0.38633297 - layer.4.v_cache 0.00000299 0.00076481 - layer.4.output 0.00016397 0.04219525 - ------------------------------------------------------------------------------------- - TOTAL 0.00302225 0.42366973 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 423004 -BPFP 2.7771 bits/point -EBPFP 5.5541 equivalent bits/point -MSE 0.423670 ----------------------- -------------------------------------------------------- -Time: 3.107s Load: 0.006s, Pack+Encode: 1.751s, Decode+Unpack: 1.350s ----------------------- -------------------------------------------------------- -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 0.4237 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample154-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample154-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample155-layer4-item1.zst (34/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample155-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: 0.005s - ------------------------------------------------------------- -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: 12,784B, BPFP=1.1480 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,388B, BPFP=3.6268 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,608B, BPFP=2.1200 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,740B, BPFP=3.7482 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,284B, BPFP=2.6297 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,296B, BPFP=3.8879 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,580B, BPFP=2.7460 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,136B, BPFP=3.7838 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,652B, BPFP=2.0341 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,064B, BPFP=3.8671 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,808B, BPFP=2.3305 -⌛️ [2/4] FRONTEND: Frontend time: 1.790s (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: 1.344s - -[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 0.03126283 4.73896526 - layer.0.v_cache 0.00000028 0.00006395 - layer.1.k_cache 0.00362225 0.32797004 - layer.1.v_cache 0.00000080 0.00021054 - layer.2.k_cache 0.00113278 0.16897786 - layer.2.v_cache 0.00000108 0.00029448 - layer.3.k_cache 0.00132193 0.19144602 - layer.3.v_cache 0.00000209 0.00048420 - layer.4.k_cache 0.00322631 0.39905180 - layer.4.v_cache 0.00000306 0.00075910 - layer.4.output 0.00017090 0.03988368 - ------------------------------------------------------------------------------------- - TOTAL 0.00294693 0.42769700 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 433340 -BPFP 2.7795 bits/point -EBPFP 5.5591 equivalent bits/point -MSE 0.427697 ----------------------- -------------------------------------------------------- -Time: 3.140s Load: 0.005s, Pack+Encode: 1.790s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -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 0.4277 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample155-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample155-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample156-layer4-item1.zst (35/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample156-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 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, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,820B, BPFP=1.1381 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,368B, BPFP=3.5838 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,880B, BPFP=2.1200 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,008B, BPFP=3.8182 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,712B, BPFP=2.5490 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,260B, BPFP=3.8406 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,996B, BPFP=2.7518 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,900B, BPFP=3.7198 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,964B, BPFP=2.0387 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,404B, BPFP=3.8533 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,564B, BPFP=2.3430 -⌛️ [2/4] FRONTEND: Frontend time: 1.795s (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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.350s - -[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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03162268 4.43531244 - layer.0.v_cache 0.00000028 0.00006411 - layer.1.k_cache 0.00348150 0.35728231 - layer.1.v_cache 0.00000080 0.00020735 - layer.2.k_cache 0.00116795 0.16633836 - layer.2.v_cache 0.00000108 0.00029750 - layer.3.k_cache 0.00133447 0.19732285 - layer.3.v_cache 0.00000211 0.00050810 - layer.4.k_cache 0.00331899 0.40335673 - layer.4.v_cache 0.00000308 0.00077989 - layer.4.output 0.00016829 0.03255788 - ------------------------------------------------------------------------------------- - TOTAL 0.00297187 0.40655008 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 436876 -BPFP 2.7704 bits/point -EBPFP 5.5407 equivalent bits/point -MSE 0.406550 ----------------------- -------------------------------------------------------- -Time: 3.152s Load: 0.006s, Pack+Encode: 1.795s, Decode+Unpack: 1.350s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4066 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample156-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample156-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample157-layer4-item1.zst (36/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample157-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: 0.006s - ------------------------------------------------------------- -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: 12,984B, BPFP=1.1271 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,416B, BPFP=3.5951 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,848B, BPFP=2.0701 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,944B, BPFP=3.8146 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,520B, BPFP=2.6493 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,068B, BPFP=3.8253 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,188B, BPFP=2.7941 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,056B, BPFP=3.7375 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,140B, BPFP=2.0955 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,984B, BPFP=3.9049 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,732B, BPFP=2.2728 -⌛️ [2/4] FRONTEND: Frontend time: 1.803s (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: 1.345s - -[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 0.03130752 4.59179755 - layer.0.v_cache 0.00000027 0.00006326 - layer.1.k_cache 0.00339397 0.34656639 - layer.1.v_cache 0.00000079 0.00019862 - layer.2.k_cache 0.00115273 0.17263701 - layer.2.v_cache 0.00000105 0.00027364 - layer.3.k_cache 0.00133725 0.20077748 - layer.3.v_cache 0.00000203 0.00045903 - layer.4.k_cache 0.00329241 0.43287294 - layer.4.v_cache 0.00000297 0.00076724 - layer.4.output 0.00017286 0.04782320 - ------------------------------------------------------------------------------------- - TOTAL 0.00294160 0.42412185 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 445880 -BPFP 2.7646 bits/point -EBPFP 5.5293 equivalent bits/point -MSE 0.424122 ----------------------- -------------------------------------------------------- -Time: 3.154s Load: 0.006s, Pack+Encode: 1.803s, Decode+Unpack: 1.345s ----------------------- -------------------------------------------------------- -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 0.4241 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample157-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample157-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample16-layer4-item1.zst (37/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample16-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: 0.007s - ------------------------------------------------------------- -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: 13,036B, BPFP=1.1316 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,268B, BPFP=3.5823 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,800B, BPFP=2.0660 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,648B, BPFP=3.7889 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,796B, BPFP=2.6733 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,144B, BPFP=3.8319 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,092B, BPFP=2.7858 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,496B, BPFP=3.6889 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,100B, BPFP=2.0920 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,888B, BPFP=3.8965 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,148B, BPFP=2.3036 -⌛️ [2/4] FRONTEND: Frontend time: 1.829s (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: 1.375s - -[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, 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 0.03118540 4.62341105 - layer.0.v_cache 0.00000028 0.00006225 - layer.1.k_cache 0.00362682 0.35274955 - layer.1.v_cache 0.00000082 0.00020797 - layer.2.k_cache 0.00111630 0.17146596 - layer.2.v_cache 0.00000105 0.00028011 - layer.3.k_cache 0.00134410 0.20059789 - layer.3.v_cache 0.00000205 0.00046065 - layer.4.k_cache 0.00329490 0.39900835 - layer.4.v_cache 0.00000297 0.00075448 - layer.4.output 0.00020247 0.04705957 - ------------------------------------------------------------------------------------- - TOTAL 0.00295604 0.42408833 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 446416 -BPFP 2.7680 bits/point -EBPFP 5.5359 equivalent bits/point -MSE 0.424088 ----------------------- -------------------------------------------------------- -Time: 3.211s Load: 0.007s, Pack+Encode: 1.829s, Decode+Unpack: 1.375s ----------------------- -------------------------------------------------------- -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 0.4241 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample16-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample16-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample17-layer4-item1.zst (38/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample17-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: 0.007s - ------------------------------------------------------------- -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: 13,164B, BPFP=1.1179 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,448B, BPFP=3.5197 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,544B, BPFP=2.1692 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,616B, BPFP=3.7887 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,168B, BPFP=2.6467 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,676B, BPFP=3.7938 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,776B, BPFP=2.7833 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,960B, BPFP=3.7330 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,204B, BPFP=2.0554 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,584B, BPFP=3.8709 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,768B, BPFP=2.2879 -⌛️ [2/4] FRONTEND: Frontend time: 1.820s (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: 1.350s - -[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 0.03142573 4.52942757 - layer.0.v_cache 0.00000027 0.00006279 - layer.1.k_cache 0.00360029 0.35861260 - layer.1.v_cache 0.00000081 0.00020379 - layer.2.k_cache 0.00114206 0.17009810 - layer.2.v_cache 0.00000107 0.00027491 - layer.3.k_cache 0.00131314 0.19725895 - layer.3.v_cache 0.00000205 0.00048011 - layer.4.k_cache 0.00327277 0.40649696 - layer.4.v_cache 0.00000301 0.00074712 - layer.4.output 0.00017705 0.04191191 - ------------------------------------------------------------------------------------- - TOTAL 0.00296210 0.41652218 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 454908 -BPFP 2.7593 bits/point -EBPFP 5.5186 equivalent bits/point -MSE 0.416522 ----------------------- -------------------------------------------------------- -Time: 3.177s Load: 0.007s, Pack+Encode: 1.820s, Decode+Unpack: 1.350s ----------------------- -------------------------------------------------------- -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 0.4165 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample17-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample17-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample18-layer4-item1.zst (39/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample18-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: 0.007s - ------------------------------------------------------------- -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: 12,992B, BPFP=1.1033 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,652B, BPFP=3.5370 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,796B, BPFP=2.1056 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,008B, BPFP=3.8220 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,104B, BPFP=2.6413 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,876B, BPFP=3.8108 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,600B, BPFP=2.7683 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,164B, BPFP=3.7503 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,804B, BPFP=2.0214 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,348B, BPFP=3.9358 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,740B, BPFP=2.2660 -⌛️ [2/4] FRONTEND: Frontend time: 1.816s (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: 1.345s - -[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 0.03108079 4.36027494 - layer.0.v_cache 0.00000027 0.00006339 - layer.1.k_cache 0.00341615 0.33705915 - layer.1.v_cache 0.00000080 0.00019546 - layer.2.k_cache 0.00113779 0.16857413 - layer.2.v_cache 0.00000105 0.00026479 - layer.3.k_cache 0.00132123 0.19857614 - layer.3.v_cache 0.00000201 0.00044088 - layer.4.k_cache 0.00323701 0.39328554 - layer.4.v_cache 0.00000300 0.00074159 - layer.4.output 0.00016915 0.04915056 - ------------------------------------------------------------------------------------- - TOTAL 0.00291977 0.40400559 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 454084 -BPFP 2.7543 bits/point -EBPFP 5.5086 equivalent bits/point -MSE 0.404006 ----------------------- -------------------------------------------------------- -Time: 3.168s Load: 0.007s, Pack+Encode: 1.816s, Decode+Unpack: 1.345s ----------------------- -------------------------------------------------------- -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 0.4040 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample18-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample18-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample19-layer4-item1.zst (40/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample19-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: 0.006s - ------------------------------------------------------------- -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: 13,036B, BPFP=1.1316 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,316B, BPFP=3.5865 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,216B, BPFP=2.1021 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,544B, BPFP=3.7799 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,244B, BPFP=2.6253 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,120B, BPFP=3.8299 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,144B, BPFP=2.7903 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,504B, BPFP=3.6896 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,512B, BPFP=2.1278 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,556B, BPFP=3.8677 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,180B, BPFP=2.3043 -⌛️ [2/4] FRONTEND: Frontend time: 1.767s (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: 1.348s - -[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 0.03108226 4.51276788 - layer.0.v_cache 0.00000028 0.00006176 - layer.1.k_cache 0.00355800 0.35760422 - layer.1.v_cache 0.00000082 0.00020727 - layer.2.k_cache 0.00111789 0.17067479 - layer.2.v_cache 0.00000106 0.00027807 - layer.3.k_cache 0.00134609 0.20648496 - layer.3.v_cache 0.00000204 0.00045789 - layer.4.k_cache 0.00329523 0.42408947 - layer.4.v_cache 0.00000295 0.00076211 - layer.4.output 0.00029330 0.05244524 - ------------------------------------------------------------------------------------- - TOTAL 0.00296998 0.42022638 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 446372 -BPFP 2.7677 bits/point -EBPFP 5.5354 equivalent bits/point -MSE 0.420226 ----------------------- -------------------------------------------------------- -Time: 3.121s Load: 0.006s, Pack+Encode: 1.767s, Decode+Unpack: 1.348s ----------------------- -------------------------------------------------------- -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 0.4202 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample19-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample19-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample2-layer4-item1.zst (41/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample2-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: 13,756B, BPFP=1.0966 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 45,080B, BPFP=3.5938 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,972B, BPFP=2.2299 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,120B, BPFP=3.6767 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,816B, BPFP=2.6958 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,480B, BPFP=3.7054 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,108B, BPFP=2.7191 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 46,080B, BPFP=3.6735 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 28,836B, BPFP=2.2988 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,024B, BPFP=3.7487 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 119,788B, BPFP=2.3874 -⌛️ [2/4] FRONTEND: Frontend time: 1.813s (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: 1.344s - -[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 0.03050847 4.89748834 - layer.0.v_cache 0.00000027 0.00006131 - layer.1.k_cache 0.00347784 0.37109663 - layer.1.v_cache 0.00000091 0.00019735 - layer.2.k_cache 0.00115394 0.17126586 - layer.2.v_cache 0.00000107 0.00027168 - layer.3.k_cache 0.00132167 0.19880410 - layer.3.v_cache 0.00000207 0.00046232 - layer.4.k_cache 0.00336398 0.47440498 - layer.4.v_cache 0.00000299 0.00074007 - layer.4.output 0.00016871 0.04231373 - ------------------------------------------------------------------------------------- - TOTAL 0.00289343 0.44886054 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 489060 -BPFP 2.7848 bits/point -EBPFP 5.5697 equivalent bits/point -MSE 0.448861 ----------------------- -------------------------------------------------------- -Time: 3.164s Load: 0.007s, Pack+Encode: 1.813s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -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 0.4489 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample2-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample2-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample20-layer4-item1.zst (42/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample20-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: 13,072B, BPFP=1.0981 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,448B, BPFP=3.4819 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,676B, BPFP=2.1569 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,004B, BPFP=3.7806 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,328B, BPFP=2.6317 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,756B, BPFP=3.8438 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,328B, BPFP=2.7157 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,984B, BPFP=3.6949 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,404B, BPFP=2.1341 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,108B, BPFP=3.8733 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,004B, BPFP=2.2472 -⌛️ [2/4] FRONTEND: Frontend time: 1.769s (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: 1.351s - -[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 0.03217121 4.27351954 - layer.0.v_cache 0.00000028 0.00006243 - layer.1.k_cache 0.00347451 0.33788583 - layer.1.v_cache 0.00000083 0.00019331 - layer.2.k_cache 0.00115003 0.16638594 - layer.2.v_cache 0.00000108 0.00027071 - layer.3.k_cache 0.00134333 0.19426651 - layer.3.v_cache 0.00000208 0.00045571 - layer.4.k_cache 0.00325638 0.41161855 - layer.4.v_cache 0.00000311 0.00077279 - layer.4.output 0.00016214 0.04855244 - ------------------------------------------------------------------------------------- - TOTAL 0.00300367 0.39854579 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 457112 -BPFP 2.7428 bits/point -EBPFP 5.4857 equivalent bits/point -MSE 0.398546 ----------------------- -------------------------------------------------------- -Time: 3.127s Load: 0.006s, Pack+Encode: 1.769s, Decode+Unpack: 1.351s ----------------------- -------------------------------------------------------- -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 0.3985 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample20-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample20-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample21-layer4-item1.zst (43/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample21-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: 12,936B, BPFP=1.0751 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,400B, BPFP=3.4408 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,640B, BPFP=2.1310 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,872B, BPFP=3.7294 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,336B, BPFP=2.6875 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,456B, BPFP=3.7779 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,868B, BPFP=2.7317 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,036B, BPFP=3.6599 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,924B, BPFP=2.0715 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,676B, BPFP=3.7962 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,588B, BPFP=2.1731 -⌛️ [2/4] FRONTEND: Frontend time: 1.794s (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: 1.356s - -[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 0.03064084 4.42862133 - layer.0.v_cache 0.00000027 0.00005927 - layer.1.k_cache 0.00361138 0.36351934 - layer.1.v_cache 0.00000084 0.00019141 - layer.2.k_cache 0.00114431 0.17540952 - layer.2.v_cache 0.00000106 0.00026863 - layer.3.k_cache 0.00136507 0.19873262 - layer.3.v_cache 0.00000197 0.00043677 - layer.4.k_cache 0.00326917 0.44150973 - layer.4.v_cache 0.00000291 0.00071874 - layer.4.output 0.00019732 0.04938997 - ------------------------------------------------------------------------------------- - TOTAL 0.00291622 0.41478766 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 454732 -BPFP 2.6995 bits/point -EBPFP 5.3991 equivalent bits/point -MSE 0.414788 ----------------------- -------------------------------------------------------- -Time: 3.157s Load: 0.007s, Pack+Encode: 1.794s, Decode+Unpack: 1.356s ----------------------- -------------------------------------------------------- -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 0.4148 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample21-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample21-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample23-layer4-item1.zst (44/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample23-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: 13,468B, BPFP=1.0960 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 43,368B, BPFP=3.5293 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,704B, BPFP=2.1732 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,492B, BPFP=3.7835 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,652B, BPFP=2.6572 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,304B, BPFP=3.7682 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,240B, BPFP=2.7051 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,664B, BPFP=3.7161 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,088B, BPFP=2.1230 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,368B, BPFP=3.7734 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 113,568B, BPFP=2.3105 -⌛️ [2/4] FRONTEND: Frontend time: 1.842s (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: 1.366s - -[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 0.03181871 4.68024381 - layer.0.v_cache 0.00000028 0.00006361 - layer.1.k_cache 0.00341635 0.35507210 - layer.1.v_cache 0.00000081 0.00019973 - layer.2.k_cache 0.00112215 0.17228651 - layer.2.v_cache 0.00000111 0.00028114 - layer.3.k_cache 0.00131706 0.20067743 - layer.3.v_cache 0.00000210 0.00046903 - layer.4.k_cache 0.00330853 0.42958967 - layer.4.v_cache 0.00000300 0.00073138 - layer.4.output 0.00017506 0.03688100 - ------------------------------------------------------------------------------------- - TOTAL 0.00297788 0.42765274 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 473916 -BPFP 2.7548 bits/point -EBPFP 5.5096 equivalent bits/point -MSE 0.427653 ----------------------- -------------------------------------------------------- -Time: 3.214s Load: 0.007s, Pack+Encode: 1.842s, Decode+Unpack: 1.366s ----------------------- -------------------------------------------------------- -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 0.4277 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample23-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample23-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample24-layer4-item1.zst (45/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample24-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: 13,012B, BPFP=1.1050 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,520B, BPFP=3.5258 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,728B, BPFP=2.0999 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,796B, BPFP=3.8040 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,288B, BPFP=2.6569 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,808B, BPFP=3.8050 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,400B, BPFP=2.7514 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,512B, BPFP=3.7799 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,648B, BPFP=2.0931 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,364B, BPFP=3.9372 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,288B, BPFP=2.2777 -⌛️ [2/4] FRONTEND: Frontend time: 1.789s (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: 1.327s - -[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 0.03098850 4.25903752 - layer.0.v_cache 0.00000027 0.00006282 - layer.1.k_cache 0.00332311 0.32831555 - layer.1.v_cache 0.00000079 0.00019335 - layer.2.k_cache 0.00114598 0.16832459 - layer.2.v_cache 0.00000105 0.00026799 - layer.3.k_cache 0.00130880 0.19604393 - layer.3.v_cache 0.00000204 0.00044671 - layer.4.k_cache 0.00318646 0.41177148 - layer.4.v_cache 0.00000304 0.00074336 - layer.4.output 0.00017235 0.04529334 - ------------------------------------------------------------------------------------- - TOTAL 0.00290353 0.39617005 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 455364 -BPFP 2.7621 bits/point -EBPFP 5.5241 equivalent bits/point -MSE 0.396170 ----------------------- -------------------------------------------------------- -Time: 3.122s Load: 0.006s, Pack+Encode: 1.789s, Decode+Unpack: 1.327s ----------------------- -------------------------------------------------------- -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 0.3962 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample24-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample24-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample27-layer4-item1.zst (46/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample27-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: 13,332B, BPFP=1.0964 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,272B, BPFP=3.4763 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,996B, BPFP=2.1378 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,132B, BPFP=3.7115 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,056B, BPFP=2.6362 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,400B, BPFP=3.7336 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,888B, BPFP=2.7046 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,700B, BPFP=3.6760 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,844B, BPFP=2.1253 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,736B, BPFP=3.7612 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 109,288B, BPFP=2.2469 -⌛️ [2/4] FRONTEND: Frontend time: 1.761s (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: 1.345s - -[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 0.03093621 4.57803666 - layer.0.v_cache 0.00000027 0.00006416 - layer.1.k_cache 0.00341424 0.35365279 - layer.1.v_cache 0.00000077 0.00019065 - layer.2.k_cache 0.00114464 0.17256116 - layer.2.v_cache 0.00000103 0.00027093 - layer.3.k_cache 0.00132643 0.19330996 - layer.3.v_cache 0.00000197 0.00044887 - layer.4.k_cache 0.00330159 0.42774610 - layer.4.v_cache 0.00000302 0.00077085 - layer.4.output 0.00016139 0.03428654 - ------------------------------------------------------------------------------------- - TOTAL 0.00291255 0.41887131 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 462644 -BPFP 2.7176 bits/point -EBPFP 5.4352 equivalent bits/point -MSE 0.418871 ----------------------- -------------------------------------------------------- -Time: 3.114s Load: 0.008s, Pack+Encode: 1.761s, Decode+Unpack: 1.345s ----------------------- -------------------------------------------------------- -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 0.4189 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample27-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample27-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample28-layer4-item1.zst (47/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample28-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: 0.007s - ------------------------------------------------------------- -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: 13,276B, BPFP=1.0918 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,468B, BPFP=3.4924 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,800B, BPFP=2.1217 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,960B, BPFP=3.6974 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,240B, BPFP=2.6513 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,528B, BPFP=3.7441 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,852B, BPFP=2.7016 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,692B, BPFP=3.6753 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,576B, BPFP=2.1033 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,048B, BPFP=3.7868 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 109,228B, BPFP=2.2456 -⌛️ [2/4] FRONTEND: Frontend time: 1.818s (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: 1.387s - -[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 0.03089984 4.48133930 - layer.0.v_cache 0.00000027 0.00006260 - layer.1.k_cache 0.00345231 0.32370290 - layer.1.v_cache 0.00000077 0.00018909 - layer.2.k_cache 0.00115012 0.17142053 - layer.2.v_cache 0.00000103 0.00026885 - layer.3.k_cache 0.00131461 0.19247115 - layer.3.v_cache 0.00000196 0.00044917 - layer.4.k_cache 0.00336511 0.42944649 - layer.4.v_cache 0.00000303 0.00077050 - layer.4.output 0.00016181 0.03378492 - ------------------------------------------------------------------------------------- - TOTAL 0.00291688 0.40966145 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 462668 -BPFP 2.7177 bits/point -EBPFP 5.4355 equivalent bits/point -MSE 0.409661 ----------------------- -------------------------------------------------------- -Time: 3.211s Load: 0.007s, Pack+Encode: 1.818s, Decode+Unpack: 1.387s ----------------------- -------------------------------------------------------- -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 0.4097 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample28-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample28-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample29-layer4-item1.zst (48/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample29-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: 0.007s - ------------------------------------------------------------- -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: 12,900B, BPFP=1.0954 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,756B, BPFP=3.5459 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,876B, BPFP=2.1124 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,672B, BPFP=3.7935 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,944B, BPFP=2.6277 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,656B, BPFP=3.7921 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,580B, BPFP=2.7666 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,660B, BPFP=3.7925 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,712B, BPFP=2.0985 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,712B, BPFP=3.9667 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,036B, BPFP=2.2299 -⌛️ [2/4] FRONTEND: Frontend time: 1.799s (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: 1.344s - -[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 0.03058780 4.26994490 - layer.0.v_cache 0.00000028 0.00006500 - layer.1.k_cache 0.00355631 0.36163513 - layer.1.v_cache 0.00000081 0.00020527 - layer.2.k_cache 0.00116497 0.16985950 - layer.2.v_cache 0.00000107 0.00027676 - layer.3.k_cache 0.00134281 0.19714656 - layer.3.v_cache 0.00000206 0.00046296 - layer.4.k_cache 0.00330740 0.41428259 - layer.4.v_cache 0.00000300 0.00072782 - layer.4.output 0.00016528 0.04640059 - ------------------------------------------------------------------------------------- - TOTAL 0.00290197 0.40001492 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 453504 -BPFP 2.7508 bits/point -EBPFP 5.5016 equivalent bits/point -MSE 0.400015 ----------------------- -------------------------------------------------------- -Time: 3.150s Load: 0.007s, Pack+Encode: 1.799s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -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 0.4000 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample29-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample29-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample30-layer4-item1.zst (49/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample30-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: 13,344B, BPFP=1.0638 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 44,624B, BPFP=3.5574 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 28,152B, BPFP=2.2443 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,592B, BPFP=3.7143 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,864B, BPFP=2.6996 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 47,168B, BPFP=3.7602 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 34,260B, BPFP=2.7312 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 46,184B, BPFP=3.6818 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 29,320B, BPFP=2.3374 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,212B, BPFP=3.7637 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 115,656B, BPFP=2.3050 -⌛️ [2/4] FRONTEND: Frontend time: 1.826s (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: 1.348s - -[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 0.02931073 4.59274603 - layer.0.v_cache 0.00000028 0.00006276 - layer.1.k_cache 0.00346088 0.37461779 - layer.1.v_cache 0.00000093 0.00019359 - layer.2.k_cache 0.00114617 0.17224360 - layer.2.v_cache 0.00000108 0.00027021 - layer.3.k_cache 0.00130775 0.19161279 - layer.3.v_cache 0.00000199 0.00043924 - layer.4.k_cache 0.00329454 0.46643148 - layer.4.v_cache 0.00000306 0.00073824 - layer.4.output 0.00016599 0.04215126 - ------------------------------------------------------------------------------------- - TOTAL 0.00279938 0.42628291 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 486376 -BPFP 2.7695 bits/point -EBPFP 5.5391 equivalent bits/point -MSE 0.426283 ----------------------- -------------------------------------------------------- -Time: 3.180s Load: 0.006s, Pack+Encode: 1.826s, Decode+Unpack: 1.348s ----------------------- -------------------------------------------------------- -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 0.4263 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample30-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample30-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample31-layer4-item1.zst (50/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample31-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: 13,172B, BPFP=1.0719 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 43,116B, BPFP=3.5088 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,316B, BPFP=2.1416 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,508B, BPFP=3.7035 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,640B, BPFP=2.6562 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,632B, BPFP=3.7949 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,216B, BPFP=2.7031 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,688B, BPFP=3.7181 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,376B, BPFP=2.1465 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,252B, BPFP=3.7640 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 109,760B, BPFP=2.2331 -⌛️ [2/4] FRONTEND: Frontend time: 1.788s (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: 1.335s - -[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 0.03011865 4.57022985 - layer.0.v_cache 0.00000028 0.00005956 - layer.1.k_cache 0.00341880 0.37867141 - layer.1.v_cache 0.00000077 0.00018629 - layer.2.k_cache 0.00113755 0.17824874 - layer.2.v_cache 0.00000103 0.00026512 - layer.3.k_cache 0.00137011 0.20528332 - layer.3.v_cache 0.00000200 0.00044647 - layer.4.k_cache 0.00324136 0.44635272 - layer.4.v_cache 0.00000292 0.00074358 - layer.4.output 0.00022292 0.04258633 - ------------------------------------------------------------------------------------- - TOTAL 0.00287037 0.42505946 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 468676 -BPFP 2.7244 bits/point -EBPFP 5.4487 equivalent bits/point -MSE 0.425059 ----------------------- -------------------------------------------------------- -Time: 3.130s Load: 0.007s, Pack+Encode: 1.788s, Decode+Unpack: 1.335s ----------------------- -------------------------------------------------------- -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 0.4251 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample31-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample31-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample32-layer4-item1.zst (51/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample32-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: 13,064B, BPFP=1.0858 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,480B, BPFP=3.4475 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,048B, BPFP=2.1649 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,120B, BPFP=3.6669 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,892B, BPFP=2.6506 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,048B, BPFP=3.7440 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,924B, BPFP=2.7364 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,036B, BPFP=3.6599 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,984B, BPFP=2.0765 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,644B, BPFP=3.7936 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,948B, BPFP=2.2014 -⌛️ [2/4] FRONTEND: Frontend time: 1.790s (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: 1.344s - -[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 0.03206433 3.97223615 - layer.0.v_cache 0.00000027 0.00006288 - layer.1.k_cache 0.00350816 0.37023017 - layer.1.v_cache 0.00000080 0.00020094 - layer.2.k_cache 0.00113886 0.16995491 - layer.2.v_cache 0.00000104 0.00026845 - layer.3.k_cache 0.00133101 0.19559262 - layer.3.v_cache 0.00000202 0.00046565 - layer.4.k_cache 0.00327566 0.41857849 - layer.4.v_cache 0.00000298 0.00075388 - layer.4.output 0.00018563 0.04941616 - ------------------------------------------------------------------------------------- - TOTAL 0.00300483 0.38042920 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 455188 -BPFP 2.7022 bits/point -EBPFP 5.4045 equivalent bits/point -MSE 0.380429 ----------------------- -------------------------------------------------------- -Time: 3.140s Load: 0.007s, Pack+Encode: 1.790s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -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 0.3804 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample32-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample32-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample33-layer4-item1.zst (52/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample33-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: 13,296B, BPFP=1.0709 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 43,048B, BPFP=3.4671 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,776B, BPFP=2.1566 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,184B, BPFP=3.7197 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,772B, BPFP=2.6395 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,864B, BPFP=3.6939 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,300B, BPFP=2.6820 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,820B, BPFP=3.6904 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,352B, BPFP=2.2030 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,656B, BPFP=3.8383 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 114,272B, BPFP=2.3009 -⌛️ [2/4] FRONTEND: Frontend time: 1.747s (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: 1.352s - -[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 0.02997011 4.63368461 - layer.0.v_cache 0.00000027 0.00005876 - layer.1.k_cache 0.00344926 0.36784316 - layer.1.v_cache 0.00000085 0.00019372 - layer.2.k_cache 0.00114066 0.17281261 - layer.2.v_cache 0.00000106 0.00027061 - layer.3.k_cache 0.00133063 0.19888919 - layer.3.v_cache 0.00000201 0.00044370 - layer.4.k_cache 0.00322564 0.45920696 - layer.4.v_cache 0.00000314 0.00076369 - layer.4.output 0.00016478 0.04074602 - ------------------------------------------------------------------------------------- - TOTAL 0.00284162 0.42836794 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 476340 -BPFP 2.7404 bits/point -EBPFP 5.4807 equivalent bits/point -MSE 0.428368 ----------------------- -------------------------------------------------------- -Time: 3.106s Load: 0.007s, Pack+Encode: 1.747s, Decode+Unpack: 1.352s ----------------------- -------------------------------------------------------- -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 0.4284 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample33-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample33-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample35-layer4-item1.zst (53/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample35-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: 0.007s - ------------------------------------------------------------- -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: 13,216B, BPFP=1.0868 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,308B, BPFP=3.4793 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,852B, BPFP=2.1260 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,464B, BPFP=3.7388 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,052B, BPFP=2.6359 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,672B, BPFP=3.7559 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,876B, BPFP=2.7036 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,724B, BPFP=3.6780 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,500B, BPFP=2.0148 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,992B, BPFP=3.7822 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,256B, BPFP=2.2051 -⌛️ [2/4] FRONTEND: Frontend time: 1.817s (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: 1.376s - -[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, 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 0.03212489 4.56008622 - layer.0.v_cache 0.00000028 0.00006498 - layer.1.k_cache 0.00341497 0.35851264 - layer.1.v_cache 0.00000078 0.00020133 - layer.2.k_cache 0.00114913 0.17214518 - layer.2.v_cache 0.00000106 0.00028214 - layer.3.k_cache 0.00131997 0.19345097 - layer.3.v_cache 0.00000202 0.00044489 - layer.4.k_cache 0.00329406 0.39857045 - layer.4.v_cache 0.00000310 0.00076590 - layer.4.output 0.00015936 0.03376855 - ------------------------------------------------------------------------------------- - TOTAL 0.00299627 0.41568564 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 459912 -BPFP 2.7016 bits/point -EBPFP 5.4031 equivalent bits/point -MSE 0.415686 ----------------------- -------------------------------------------------------- -Time: 3.201s Load: 0.007s, Pack+Encode: 1.817s, Decode+Unpack: 1.376s ----------------------- -------------------------------------------------------- -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 0.4157 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample35-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample35-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample36-layer4-item1.zst (54/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample36-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: 12,924B, BPFP=1.0975 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,512B, BPFP=3.5251 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,872B, BPFP=2.1121 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,836B, BPFP=3.8074 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,560B, BPFP=2.5951 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,996B, BPFP=3.8210 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,276B, BPFP=2.7408 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,148B, BPFP=3.7490 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,732B, BPFP=2.0153 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,488B, BPFP=3.8628 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,388B, BPFP=2.2798 -⌛️ [2/4] FRONTEND: Frontend time: 1.760s (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: 1.360s - -[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 0.03130991 4.44503187 - layer.0.v_cache 0.00000027 0.00006380 - layer.1.k_cache 0.00358715 0.34571069 - layer.1.v_cache 0.00000080 0.00020334 - layer.2.k_cache 0.00113844 0.16801148 - layer.2.v_cache 0.00000105 0.00028117 - layer.3.k_cache 0.00133437 0.19911096 - layer.3.v_cache 0.00000205 0.00048148 - layer.4.k_cache 0.00328807 0.39578674 - layer.4.v_cache 0.00000299 0.00075744 - layer.4.output 0.00019541 0.04124752 - ------------------------------------------------------------------------------------- - TOTAL 0.00296048 0.40860208 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 452732 -BPFP 2.7461 bits/point -EBPFP 5.4922 equivalent bits/point -MSE 0.408602 ----------------------- -------------------------------------------------------- -Time: 3.126s Load: 0.005s, Pack+Encode: 1.760s, Decode+Unpack: 1.360s ----------------------- -------------------------------------------------------- -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 0.4086 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample36-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample36-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample37-layer4-item1.zst (55/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample37-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: 0.007s - ------------------------------------------------------------- -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: 13,100B, BPFP=1.0773 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,708B, BPFP=3.5122 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,664B, BPFP=2.1928 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,344B, BPFP=3.7289 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,488B, BPFP=2.6717 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,524B, BPFP=3.7437 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,132B, BPFP=2.7247 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,316B, BPFP=3.6444 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,380B, BPFP=2.0872 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,812B, BPFP=3.7674 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,944B, BPFP=2.2192 -⌛️ [2/4] FRONTEND: Frontend time: 1.814s (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: 1.360s - -[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 0.03212384 4.42923134 - layer.0.v_cache 0.00000028 0.00006171 - layer.1.k_cache 0.00332589 0.34184948 - layer.1.v_cache 0.00000080 0.00020045 - layer.2.k_cache 0.00116734 0.17293864 - layer.2.v_cache 0.00000106 0.00027371 - layer.3.k_cache 0.00135603 0.19959765 - layer.3.v_cache 0.00000201 0.00045009 - layer.4.k_cache 0.00321713 0.42826341 - layer.4.v_cache 0.00000315 0.00077507 - layer.4.output 0.00019059 0.04656307 - ------------------------------------------------------------------------------------- - TOTAL 0.00299713 0.41142099 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 462412 -BPFP 2.7162 bits/point -EBPFP 5.4325 equivalent bits/point -MSE 0.411421 ----------------------- -------------------------------------------------------- -Time: 3.181s Load: 0.007s, Pack+Encode: 1.814s, Decode+Unpack: 1.360s ----------------------- -------------------------------------------------------- -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 0.4114 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample37-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample37-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample39-layer4-item1.zst (56/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample39-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: 0.009s - ------------------------------------------------------------- -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: 13,032B, BPFP=1.0948 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,924B, BPFP=3.5218 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,636B, BPFP=2.1536 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,080B, BPFP=3.7870 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,472B, BPFP=2.6438 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,244B, BPFP=3.8007 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,736B, BPFP=2.7500 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,288B, BPFP=3.7204 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,052B, BPFP=2.0205 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,640B, BPFP=3.8340 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,708B, BPFP=2.2830 -⌛️ [2/4] FRONTEND: Frontend time: 1.796s (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: 1.354s - -[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 0.03063477 4.14184045 - layer.0.v_cache 0.00000027 0.00006237 - layer.1.k_cache 0.00346550 0.35439793 - layer.1.v_cache 0.00000078 0.00019691 - layer.2.k_cache 0.00113041 0.17091718 - layer.2.v_cache 0.00000112 0.00028018 - layer.3.k_cache 0.00132028 0.19941309 - layer.3.v_cache 0.00000203 0.00045905 - layer.4.k_cache 0.00322248 0.39856170 - layer.4.v_cache 0.00000299 0.00075803 - layer.4.output 0.00017253 0.04881979 - ------------------------------------------------------------------------------------- - TOTAL 0.00289077 0.39015472 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 457812 -BPFP 2.7470 bits/point -EBPFP 5.4941 equivalent bits/point -MSE 0.390155 ----------------------- -------------------------------------------------------- -Time: 3.158s Load: 0.009s, Pack+Encode: 1.796s, Decode+Unpack: 1.354s ----------------------- -------------------------------------------------------- -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 0.3902 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample39-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample39-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample4-layer4-item1.zst (57/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample4-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: 14,176B, BPFP=1.1075 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 45,476B, BPFP=3.5528 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 29,536B, BPFP=2.3075 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 47,024B, BPFP=3.6738 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,620B, BPFP=2.7047 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 47,180B, BPFP=3.6859 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 35,328B, BPFP=2.7600 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 46,236B, BPFP=3.6122 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 30,636B, BPFP=2.3934 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,804B, BPFP=3.7347 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 127,116B, BPFP=2.4827 -⌛️ [2/4] FRONTEND: Frontend time: 1.796s (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: 1.363s - -[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 0.03019520 4.97418274 - layer.0.v_cache 0.00000027 0.00006179 - layer.1.k_cache 0.00345483 0.37748795 - layer.1.v_cache 0.00000092 0.00019718 - layer.2.k_cache 0.00115188 0.16789993 - layer.2.v_cache 0.00000108 0.00026485 - layer.3.k_cache 0.00131981 0.19160950 - layer.3.v_cache 0.00000202 0.00045130 - layer.4.k_cache 0.00326509 0.45686768 - layer.4.v_cache 0.00000302 0.00073296 - layer.4.output 0.00016666 0.03653471 - ------------------------------------------------------------------------------------- - TOTAL 0.00286148 0.45113534 - (elements=1,433,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1433600 -Total Bytes 505132 -BPFP 2.8188 bits/point -EBPFP 5.6376 equivalent bits/point -MSE 0.451135 ----------------------- -------------------------------------------------------- -Time: 3.166s Load: 0.007s, Pack+Encode: 1.796s, Decode+Unpack: 1.363s ----------------------- -------------------------------------------------------- -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 0.4511 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample4-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample4-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample40-layer4-item1.zst (58/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample40-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: 13,064B, BPFP=1.1468 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,932B, BPFP=3.5930 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,540B, BPFP=2.0664 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,112B, BPFP=3.8722 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,336B, BPFP=2.5751 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,924B, BPFP=3.8557 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,584B, BPFP=2.7725 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,080B, BPFP=3.6938 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,308B, BPFP=2.0460 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,248B, BPFP=3.8841 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,468B, BPFP=2.3804 -⌛️ [2/4] FRONTEND: Frontend time: 1.804s (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: 1.360s - -[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 0.03166650 4.26413032 - layer.0.v_cache 0.00000027 0.00006129 - layer.1.k_cache 0.00343321 0.33996824 - layer.1.v_cache 0.00000094 0.00021588 - layer.2.k_cache 0.00113193 0.17063527 - layer.2.v_cache 0.00000110 0.00027998 - layer.3.k_cache 0.00132786 0.20315247 - layer.3.v_cache 0.00000214 0.00047147 - layer.4.k_cache 0.00323286 0.40478571 - layer.4.v_cache 0.00000312 0.00077627 - layer.4.output 0.00018081 0.05178430 - ------------------------------------------------------------------------------------- - TOTAL 0.00296594 0.39940101 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 444596 -BPFP 2.7876 bits/point -EBPFP 5.5753 equivalent bits/point -MSE 0.399401 ----------------------- -------------------------------------------------------- -Time: 3.170s Load: 0.006s, Pack+Encode: 1.804s, Decode+Unpack: 1.360s ----------------------- -------------------------------------------------------- -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 0.3994 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample40-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample40-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample41-layer4-item1.zst (59/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample41-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: 13,096B, BPFP=1.1001 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,204B, BPFP=3.4614 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,620B, BPFP=2.1522 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,068B, BPFP=3.7860 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,212B, BPFP=2.7060 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,940B, BPFP=3.7752 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,300B, BPFP=2.7134 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,984B, BPFP=3.6949 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,036B, BPFP=2.1872 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,084B, BPFP=3.8713 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 109,612B, BPFP=2.3020 -⌛️ [2/4] FRONTEND: Frontend time: 1.827s (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: 1.345s - -[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 0.03068307 4.31376139 - layer.0.v_cache 0.00000028 0.00006348 - layer.1.k_cache 0.00346304 0.35932307 - layer.1.v_cache 0.00000079 0.00019949 - layer.2.k_cache 0.00114171 0.16985450 - layer.2.v_cache 0.00000107 0.00026993 - layer.3.k_cache 0.00131636 0.19141540 - layer.3.v_cache 0.00000209 0.00046089 - layer.4.k_cache 0.00323274 0.42948019 - layer.4.v_cache 0.00000311 0.00075454 - layer.4.output 0.00016873 0.04623202 - ------------------------------------------------------------------------------------- - TOTAL 0.00289423 0.40360793 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 460156 -BPFP 2.7611 bits/point -EBPFP 5.5222 equivalent bits/point -MSE 0.403608 ----------------------- -------------------------------------------------------- -Time: 3.178s Load: 0.006s, Pack+Encode: 1.827s, Decode+Unpack: 1.345s ----------------------- -------------------------------------------------------- -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 0.4036 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample41-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample41-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample43-layer4-item1.zst (60/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample43-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: 0.007s - ------------------------------------------------------------- -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: 12,840B, BPFP=1.1146 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,080B, BPFP=3.5660 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,912B, BPFP=2.0757 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,796B, BPFP=3.8017 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,520B, BPFP=2.6493 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,984B, BPFP=3.9049 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,036B, BPFP=2.7809 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,620B, BPFP=3.7865 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,560B, BPFP=2.0451 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,372B, BPFP=3.9385 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,400B, BPFP=2.2873 -⌛️ [2/4] FRONTEND: Frontend time: 1.811s (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: 1.355s - -[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 0.03151998 4.53200548 - layer.0.v_cache 0.00000027 0.00006185 - layer.1.k_cache 0.00334378 0.33093260 - layer.1.v_cache 0.00000080 0.00019685 - layer.2.k_cache 0.00114309 0.17294939 - layer.2.v_cache 0.00000106 0.00027035 - layer.3.k_cache 0.00133556 0.19459856 - layer.3.v_cache 0.00000205 0.00044440 - layer.4.k_cache 0.00317137 0.38889694 - layer.4.v_cache 0.00000297 0.00075595 - layer.4.output 0.00023356 0.04591870 - ------------------------------------------------------------------------------------- - TOTAL 0.00296108 0.41462765 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 447120 -BPFP 2.7723 bits/point -EBPFP 5.5446 equivalent bits/point -MSE 0.414628 ----------------------- -------------------------------------------------------- -Time: 3.173s Load: 0.007s, Pack+Encode: 1.811s, Decode+Unpack: 1.355s ----------------------- -------------------------------------------------------- -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 0.4146 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample43-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample43-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample44-layer4-item1.zst (61/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample44-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: 12,684B, BPFP=1.1390 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,204B, BPFP=3.6103 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,688B, BPFP=2.1272 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,268B, BPFP=3.7956 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,048B, BPFP=2.6085 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,372B, BPFP=3.8948 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,140B, BPFP=2.7963 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,784B, BPFP=3.8420 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,724B, BPFP=2.0406 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,744B, BPFP=3.9282 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,424B, BPFP=2.3218 -⌛️ [2/4] FRONTEND: Frontend time: 1.834s (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: 1.374s - -[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 0.03168404 4.69022343 - layer.0.v_cache 0.00000027 0.00006274 - layer.1.k_cache 0.00359645 0.34590875 - layer.1.v_cache 0.00000085 0.00020010 - layer.2.k_cache 0.00115081 0.16575694 - layer.2.v_cache 0.00000106 0.00027872 - layer.3.k_cache 0.00133256 0.19212140 - layer.3.v_cache 0.00000205 0.00046303 - layer.4.k_cache 0.00340755 0.39588937 - layer.4.v_cache 0.00000299 0.00074437 - layer.4.output 0.00018399 0.04668771 - ------------------------------------------------------------------------------------- - TOTAL 0.00299390 0.42702855 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 435080 -BPFP 2.7907 bits/point -EBPFP 5.5814 equivalent bits/point -MSE 0.427029 ----------------------- -------------------------------------------------------- -Time: 3.214s Load: 0.006s, Pack+Encode: 1.834s, Decode+Unpack: 1.374s ----------------------- -------------------------------------------------------- -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 0.4270 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample44-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample44-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample46-layer4-item1.zst (62/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample46-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: 12,984B, BPFP=1.0566 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,888B, BPFP=3.4089 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,400B, BPFP=2.1484 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,564B, BPFP=3.7080 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,724B, BPFP=2.6631 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,256B, BPFP=3.7643 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,176B, BPFP=2.6999 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,348B, BPFP=3.6904 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,264B, BPFP=2.1374 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,864B, BPFP=3.7324 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,448B, BPFP=2.2064 -⌛️ [2/4] FRONTEND: Frontend time: 1.820s (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: 1.373s - -[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 0.03074032 4.50221697 - layer.0.v_cache 0.00000027 0.00006161 - layer.1.k_cache 0.00337797 0.37096159 - layer.1.v_cache 0.00000078 0.00018553 - layer.2.k_cache 0.00113475 0.17832907 - layer.2.v_cache 0.00000103 0.00026340 - layer.3.k_cache 0.00138182 0.19767020 - layer.3.v_cache 0.00000203 0.00044503 - layer.4.k_cache 0.00323877 0.44666497 - layer.4.v_cache 0.00000295 0.00073495 - layer.4.output 0.00017303 0.03103960 - ------------------------------------------------------------------------------------- - TOTAL 0.00289806 0.41583512 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 464916 -BPFP 2.7025 bits/point -EBPFP 5.4050 equivalent bits/point -MSE 0.415835 ----------------------- -------------------------------------------------------- -Time: 3.200s Load: 0.007s, Pack+Encode: 1.820s, Decode+Unpack: 1.373s ----------------------- -------------------------------------------------------- -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 0.4158 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample46-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample46-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample47-layer4-item1.zst (63/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample47-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: 12,624B, BPFP=1.1336 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,240B, BPFP=3.6135 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,504B, BPFP=2.1106 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,476B, BPFP=3.8143 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,096B, BPFP=2.6128 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,456B, BPFP=3.9023 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,220B, BPFP=2.8035 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,952B, BPFP=3.8570 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,804B, BPFP=2.0478 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,412B, BPFP=3.8983 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,336B, BPFP=2.3199 -⌛️ [2/4] FRONTEND: Frontend time: 1.801s (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: 1.331s - -[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 0.03178630 4.35097249 - layer.0.v_cache 0.00000027 0.00006253 - layer.1.k_cache 0.00361903 0.34738098 - layer.1.v_cache 0.00000083 0.00019937 - layer.2.k_cache 0.00113732 0.16553747 - layer.2.v_cache 0.00000106 0.00028032 - layer.3.k_cache 0.00132704 0.19747311 - layer.3.v_cache 0.00000206 0.00046189 - layer.4.k_cache 0.00332258 0.40853698 - layer.4.v_cache 0.00000303 0.00073998 - layer.4.output 0.00018372 0.04618403 - ------------------------------------------------------------------------------------- - TOTAL 0.00299531 0.40402723 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 435120 -BPFP 2.7909 bits/point -EBPFP 5.5819 equivalent bits/point -MSE 0.404027 ----------------------- -------------------------------------------------------- -Time: 3.138s Load: 0.006s, Pack+Encode: 1.801s, Decode+Unpack: 1.331s ----------------------- -------------------------------------------------------- -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 0.4040 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample47-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample47-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample48-layer4-item1.zst (64/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample48-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: 13,460B, BPFP=1.0841 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 43,652B, BPFP=3.5158 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 27,168B, BPFP=2.1881 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 46,140B, BPFP=3.7162 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,116B, BPFP=2.6672 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,240B, BPFP=3.7242 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,316B, BPFP=2.6833 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,768B, BPFP=3.6862 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,548B, BPFP=2.2188 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,900B, BPFP=3.7774 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 116,004B, BPFP=2.3358 -⌛️ [2/4] FRONTEND: Frontend time: 1.797s (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: 1.330s - -[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 0.03078698 4.30468907 - layer.0.v_cache 0.00000027 0.00006153 - layer.1.k_cache 0.00338495 0.38182540 - layer.1.v_cache 0.00000090 0.00018876 - layer.2.k_cache 0.00114245 0.17308862 - layer.2.v_cache 0.00000106 0.00026105 - layer.3.k_cache 0.00133426 0.19268079 - layer.3.v_cache 0.00000206 0.00044621 - layer.4.k_cache 0.00327800 0.44903250 - layer.4.v_cache 0.00000314 0.00074014 - layer.4.output 0.00016719 0.04588553 - ------------------------------------------------------------------------------------- - TOTAL 0.00290020 0.40618259 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 479312 -BPFP 2.7575 bits/point -EBPFP 5.5149 equivalent bits/point -MSE 0.406183 ----------------------- -------------------------------------------------------- -Time: 3.134s Load: 0.007s, Pack+Encode: 1.797s, Decode+Unpack: 1.330s ----------------------- -------------------------------------------------------- -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 0.4062 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample48-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample48-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample50-layer4-item1.zst (65/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample50-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: 0.007s - ------------------------------------------------------------- -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: 13,036B, BPFP=1.1070 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,596B, BPFP=3.5323 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,156B, BPFP=2.1362 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,172B, BPFP=3.8359 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,764B, BPFP=2.6124 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,580B, BPFP=3.8706 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,460B, BPFP=2.7565 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,336B, BPFP=3.7649 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,832B, BPFP=2.0238 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,020B, BPFP=3.9079 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,288B, BPFP=2.2352 -⌛️ [2/4] FRONTEND: Frontend time: 1.837s (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: 1.350s - -[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 0.03091334 4.62383502 - layer.0.v_cache 0.00000027 0.00006465 - layer.1.k_cache 0.00334237 0.33608051 - layer.1.v_cache 0.00000084 0.00019277 - layer.2.k_cache 0.00114123 0.16646343 - layer.2.v_cache 0.00000105 0.00026506 - layer.3.k_cache 0.00132733 0.19697836 - layer.3.v_cache 0.00000206 0.00044823 - layer.4.k_cache 0.00332026 0.39547448 - layer.4.v_cache 0.00000301 0.00076611 - layer.4.output 0.00018747 0.04593879 - ------------------------------------------------------------------------------------- - TOTAL 0.00291440 0.42173741 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 453240 -BPFP 2.7492 bits/point -EBPFP 5.4984 equivalent bits/point -MSE 0.421737 ----------------------- -------------------------------------------------------- -Time: 3.194s Load: 0.007s, Pack+Encode: 1.837s, Decode+Unpack: 1.350s ----------------------- -------------------------------------------------------- -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 0.4217 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample50-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample50-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample51-layer4-item1.zst (66/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample51-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: 13,076B, BPFP=1.0868 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,444B, BPFP=3.4445 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,496B, BPFP=2.2021 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,788B, BPFP=3.7224 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,936B, BPFP=2.6543 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,276B, BPFP=3.7630 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,080B, BPFP=2.7493 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,360B, BPFP=3.6868 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,052B, BPFP=2.1652 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,800B, BPFP=3.7234 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,828B, BPFP=2.2404 -⌛️ [2/4] FRONTEND: Frontend time: 1.805s (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: 1.196s - -[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 0.03107318 4.46032585 - layer.0.v_cache 0.00000027 0.00006208 - layer.1.k_cache 0.00348481 0.36748221 - layer.1.v_cache 0.00000081 0.00020129 - layer.2.k_cache 0.00116411 0.17304591 - layer.2.v_cache 0.00000106 0.00027465 - layer.3.k_cache 0.00133018 0.19501915 - layer.3.v_cache 0.00000204 0.00046303 - layer.4.k_cache 0.00337179 0.44107798 - layer.4.v_cache 0.00000293 0.00073585 - layer.4.output 0.00016660 0.03585731 - ------------------------------------------------------------------------------------- - TOTAL 0.00293554 0.41300838 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 459136 -BPFP 2.7257 bits/point -EBPFP 5.4514 equivalent bits/point -MSE 0.413008 ----------------------- -------------------------------------------------------- -Time: 3.007s Load: 0.006s, Pack+Encode: 1.805s, Decode+Unpack: 1.196s ----------------------- -------------------------------------------------------- -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 0.4130 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample51-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample51-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample53-layer4-item1.zst (67/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample53-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: 0.007s - ------------------------------------------------------------- -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: 12,396B, BPFP=1.1131 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,316B, BPFP=3.6203 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,536B, BPFP=2.1135 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,296B, BPFP=3.7981 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,124B, BPFP=2.5255 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,176B, BPFP=3.8772 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,688B, BPFP=2.7557 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,748B, BPFP=3.7489 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,556B, BPFP=2.0255 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,804B, BPFP=3.9335 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 101,820B, BPFP=2.2858 -⌛️ [2/4] FRONTEND: Frontend time: 1.715s (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: 1.336s - -[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 0.03151683 4.83144195 - layer.0.v_cache 0.00000028 0.00006559 - layer.1.k_cache 0.00361716 0.32454037 - layer.1.v_cache 0.00000078 0.00020369 - layer.2.k_cache 0.00116662 0.16506592 - layer.2.v_cache 0.00000103 0.00028064 - layer.3.k_cache 0.00135197 0.19601655 - layer.3.v_cache 0.00000203 0.00045370 - layer.4.k_cache 0.00327724 0.40128326 - layer.4.v_cache 0.00000296 0.00075221 - layer.4.output 0.00017041 0.04103159 - ------------------------------------------------------------------------------------- - TOTAL 0.00297275 0.43458787 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 430460 -BPFP 2.7611 bits/point -EBPFP 5.5221 equivalent bits/point -MSE 0.434588 ----------------------- -------------------------------------------------------- -Time: 3.058s Load: 0.007s, Pack+Encode: 1.715s, Decode+Unpack: 1.336s ----------------------- -------------------------------------------------------- -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 0.4346 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample53-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample53-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample54-layer4-item1.zst (68/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample54-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: 0.007s - ------------------------------------------------------------- -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: 12,860B, BPFP=1.0576 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,012B, BPFP=3.4549 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,988B, BPFP=2.1372 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,968B, BPFP=3.6158 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,156B, BPFP=2.6444 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,660B, BPFP=3.7549 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,084B, BPFP=2.7207 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,404B, BPFP=3.6516 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,576B, BPFP=2.1033 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,488B, BPFP=3.7408 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,048B, BPFP=2.2214 -⌛️ [2/4] FRONTEND: Frontend time: 1.772s (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: 1.343s - -[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 0.03080931 4.27603374 - layer.0.v_cache 0.00000028 0.00006239 - layer.1.k_cache 0.00333058 0.36269375 - layer.1.v_cache 0.00000087 0.00020393 - layer.2.k_cache 0.00113656 0.17687812 - layer.2.v_cache 0.00000104 0.00027714 - layer.3.k_cache 0.00135899 0.19553183 - layer.3.v_cache 0.00000196 0.00044761 - layer.4.k_cache 0.00332312 0.44048229 - layer.4.v_cache 0.00000307 0.00075918 - layer.4.output 0.00019584 0.04018078 - ------------------------------------------------------------------------------------- - TOTAL 0.00291065 0.40100665 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 459244 -BPFP 2.6976 bits/point -EBPFP 5.3953 equivalent bits/point -MSE 0.401007 ----------------------- -------------------------------------------------------- -Time: 3.122s Load: 0.007s, Pack+Encode: 1.772s, Decode+Unpack: 1.343s ----------------------- -------------------------------------------------------- -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 0.4010 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample54-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample54-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample55-layer4-item1.zst (69/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample55-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: 0.007s - ------------------------------------------------------------- -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: 12,596B, BPFP=1.1311 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,144B, BPFP=3.6049 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,232B, BPFP=2.0862 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,988B, BPFP=3.7705 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,336B, BPFP=2.5445 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,636B, BPFP=3.9185 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 29,920B, BPFP=2.6868 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,020B, BPFP=3.7733 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,732B, BPFP=2.0413 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,224B, BPFP=3.8815 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,344B, BPFP=2.3649 -⌛️ [2/4] FRONTEND: Frontend time: 1.747s (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: 1.357s - -[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 0.03144652 4.55710155 - layer.0.v_cache 0.00000028 0.00006471 - layer.1.k_cache 0.00348072 0.33294209 - layer.1.v_cache 0.00000083 0.00020730 - layer.2.k_cache 0.00116226 0.16493411 - layer.2.v_cache 0.00000106 0.00028567 - layer.3.k_cache 0.00133044 0.19264223 - layer.3.v_cache 0.00000208 0.00046833 - layer.4.k_cache 0.00328583 0.40263191 - layer.4.v_cache 0.00000293 0.00073215 - layer.4.output 0.00019032 0.03620032 - ------------------------------------------------------------------------------------- - TOTAL 0.00296244 0.41405795 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 433172 -BPFP 2.7785 bits/point -EBPFP 5.5569 equivalent bits/point -MSE 0.414058 ----------------------- -------------------------------------------------------- -Time: 3.111s Load: 0.007s, Pack+Encode: 1.747s, Decode+Unpack: 1.357s ----------------------- -------------------------------------------------------- -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 0.4141 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample55-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample55-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample59-layer4-item1.zst (70/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample59-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 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, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,984B, BPFP=1.1147 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,972B, BPFP=3.5175 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,628B, BPFP=2.0285 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,884B, BPFP=3.7675 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,540B, BPFP=2.6219 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,284B, BPFP=3.8019 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,744B, BPFP=2.7253 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,500B, BPFP=3.7345 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,476B, BPFP=2.0155 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,384B, BPFP=3.8963 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,332B, BPFP=2.2607 -⌛️ [2/4] FRONTEND: Frontend time: 1.789s (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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.352s - -[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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03082078 4.48734761 - layer.0.v_cache 0.00000027 0.00006500 - layer.1.k_cache 0.00345050 0.33441527 - layer.1.v_cache 0.00000081 0.00020122 - layer.2.k_cache 0.00112876 0.16651671 - layer.2.v_cache 0.00000108 0.00027255 - layer.3.k_cache 0.00131171 0.19089248 - layer.3.v_cache 0.00000203 0.00044789 - layer.4.k_cache 0.00324214 0.40031550 - layer.4.v_cache 0.00000302 0.00074249 - layer.4.output 0.00017376 0.04014785 - ------------------------------------------------------------------------------------- - TOTAL 0.00290401 0.41012915 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 445728 -BPFP 2.7333 bits/point -EBPFP 5.4666 equivalent bits/point -MSE 0.410129 ----------------------- -------------------------------------------------------- -Time: 3.148s Load: 0.007s, Pack+Encode: 1.789s, Decode+Unpack: 1.352s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4101 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample59-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample59-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample60-layer4-item1.zst (71/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample60-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: 0.007s - ------------------------------------------------------------- -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: 12,436B, BPFP=1.1167 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,464B, BPFP=3.6336 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,688B, BPFP=2.1272 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,480B, BPFP=3.8147 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,748B, BPFP=2.5815 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,160B, BPFP=3.8757 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,392B, BPFP=2.7292 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,224B, BPFP=3.7917 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,560B, BPFP=2.0259 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,548B, BPFP=3.9106 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 102,220B, BPFP=2.2948 -⌛️ [2/4] FRONTEND: Frontend time: 1.798s (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: 1.341s - -[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 0.03212153 4.43762453 - layer.0.v_cache 0.00000027 0.00006284 - layer.1.k_cache 0.00343529 0.32319404 - layer.1.v_cache 0.00000080 0.00019770 - layer.2.k_cache 0.00113462 0.16592257 - layer.2.v_cache 0.00000107 0.00027692 - layer.3.k_cache 0.00129755 0.19395070 - layer.3.v_cache 0.00000202 0.00046204 - layer.4.k_cache 0.00322320 0.38118314 - layer.4.v_cache 0.00000301 0.00075267 - layer.4.output 0.00017088 0.03191918 - ------------------------------------------------------------------------------------- - TOTAL 0.00299306 0.40223599 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 431920 -BPFP 2.7704 bits/point -EBPFP 5.5408 equivalent bits/point -MSE 0.402236 ----------------------- -------------------------------------------------------- -Time: 3.146s Load: 0.007s, Pack+Encode: 1.798s, Decode+Unpack: 1.341s ----------------------- -------------------------------------------------------- -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 0.4022 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample60-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample60-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample62-layer4-item1.zst (72/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample62-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: 0.006s - ------------------------------------------------------------- -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: 12,868B, BPFP=1.1170 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,496B, BPFP=3.6021 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,032B, BPFP=2.0861 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,304B, BPFP=3.7590 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,564B, BPFP=2.6531 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,112B, BPFP=3.8292 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,804B, BPFP=2.7608 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,604B, BPFP=3.6983 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,500B, BPFP=2.0399 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,828B, BPFP=3.8913 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,916B, BPFP=2.2551 -⌛️ [2/4] FRONTEND: Frontend time: 1.806s (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: 1.386s - -[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 0.03068415 4.46794942 - layer.0.v_cache 0.00000027 0.00006103 - layer.1.k_cache 0.00335417 0.34986835 - layer.1.v_cache 0.00000084 0.00019689 - layer.2.k_cache 0.00111816 0.17190384 - layer.2.v_cache 0.00000105 0.00027478 - layer.3.k_cache 0.00134433 0.19716786 - layer.3.v_cache 0.00000196 0.00044853 - layer.4.k_cache 0.00329790 0.41035076 - layer.4.v_cache 0.00000303 0.00075310 - layer.4.output 0.00016086 0.04292890 - ------------------------------------------------------------------------------------- - TOTAL 0.00288924 0.41219216 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 443028 -BPFP 2.7469 bits/point -EBPFP 5.4939 equivalent bits/point -MSE 0.412192 ----------------------- -------------------------------------------------------- -Time: 3.198s Load: 0.006s, Pack+Encode: 1.806s, Decode+Unpack: 1.386s ----------------------- -------------------------------------------------------- -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 0.4122 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample62-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample62-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample63-layer4-item1.zst (73/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample63-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: 0.007s - ------------------------------------------------------------- -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: 12,884B, BPFP=1.1184 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,872B, BPFP=3.5479 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,252B, BPFP=2.1052 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,068B, BPFP=3.7385 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,940B, BPFP=2.5990 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,660B, BPFP=3.8767 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,312B, BPFP=2.8049 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,664B, BPFP=3.7035 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,588B, BPFP=2.0476 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,952B, BPFP=3.9021 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,796B, BPFP=2.2742 -⌛️ [2/4] FRONTEND: Frontend time: 1.761s (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: 1.340s - -[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 0.03145831 4.73078817 - layer.0.v_cache 0.00000027 0.00006317 - layer.1.k_cache 0.00344612 0.33918321 - layer.1.v_cache 0.00000080 0.00020352 - layer.2.k_cache 0.00114283 0.16727575 - layer.2.v_cache 0.00000108 0.00027882 - layer.3.k_cache 0.00133751 0.20530809 - layer.3.v_cache 0.00000207 0.00045902 - layer.4.k_cache 0.00330099 0.40352804 - layer.4.v_cache 0.00000295 0.00077081 - layer.4.output 0.00016962 0.04363999 - ------------------------------------------------------------------------------------- - TOTAL 0.00295510 0.43017275 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 443988 -BPFP 2.7529 bits/point -EBPFP 5.5058 equivalent bits/point -MSE 0.430173 ----------------------- -------------------------------------------------------- -Time: 3.108s Load: 0.007s, Pack+Encode: 1.761s, Decode+Unpack: 1.340s ----------------------- -------------------------------------------------------- -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 0.4302 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample63-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample63-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample65-layer4-item1.zst (74/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample65-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 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, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,960B, BPFP=1.1126 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,124B, BPFP=3.5306 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,224B, BPFP=2.0797 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,048B, BPFP=3.6957 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,240B, BPFP=2.5962 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,896B, BPFP=3.7685 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,388B, BPFP=2.7806 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,992B, BPFP=3.6909 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,464B, BPFP=2.0144 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,268B, BPFP=3.8863 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,692B, BPFP=2.2899 -⌛️ [2/4] FRONTEND: Frontend time: 1.744s (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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.348s - -[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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03132481 4.28488696 - layer.0.v_cache 0.00000027 0.00006246 - layer.1.k_cache 0.00345143 0.34586301 - layer.1.v_cache 0.00000080 0.00019637 - layer.2.k_cache 0.00115294 0.16938572 - layer.2.v_cache 0.00000107 0.00027139 - layer.3.k_cache 0.00135241 0.20217481 - layer.3.v_cache 0.00000203 0.00043791 - layer.4.k_cache 0.00326810 0.38710936 - layer.4.v_cache 0.00000305 0.00073778 - layer.4.output 0.00025628 0.05303673 - ------------------------------------------------------------------------------------- - TOTAL 0.00297015 0.40023376 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 446296 -BPFP 2.7368 bits/point -EBPFP 5.4736 equivalent bits/point -MSE 0.400234 ----------------------- -------------------------------------------------------- -Time: 3.099s Load: 0.007s, Pack+Encode: 1.744s, Decode+Unpack: 1.348s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4002 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample65-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample65-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample66-layer4-item1.zst (75/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample66-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: 13,148B, BPFP=1.1045 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,400B, BPFP=3.4778 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,072B, BPFP=2.1902 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,468B, BPFP=3.8196 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,556B, BPFP=2.6509 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,420B, BPFP=3.7315 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,176B, BPFP=2.7030 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,904B, BPFP=3.6882 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,144B, BPFP=2.1122 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,152B, BPFP=3.8770 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,716B, BPFP=2.2622 -⌛️ [2/4] FRONTEND: Frontend time: 1.818s (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: 1.370s - -[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 0.03759774 4.17523259 - layer.0.v_cache 0.00000028 0.00006217 - layer.1.k_cache 0.00332062 0.34710640 - layer.1.v_cache 0.00000078 0.00019385 - layer.2.k_cache 0.00112828 0.17328923 - layer.2.v_cache 0.00000108 0.00027638 - layer.3.k_cache 0.00133686 0.19361793 - layer.3.v_cache 0.00000204 0.00044983 - layer.4.k_cache 0.00329063 0.41597149 - layer.4.v_cache 0.00000307 0.00074537 - layer.4.output 0.00016515 0.04873211 - ------------------------------------------------------------------------------------- - TOTAL 0.00338157 0.39299098 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 457156 -BPFP 2.7431 bits/point -EBPFP 5.4862 equivalent bits/point -MSE 0.392991 ----------------------- -------------------------------------------------------- -Time: 3.194s Load: 0.006s, Pack+Encode: 1.818s, Decode+Unpack: 1.370s ----------------------- -------------------------------------------------------- -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 0.3930 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample66-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample66-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample67-layer4-item1.zst (76/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample67-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: 13,004B, BPFP=1.0808 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,888B, BPFP=3.4814 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,908B, BPFP=2.1533 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,504B, BPFP=3.6988 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,252B, BPFP=2.6805 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,040B, BPFP=3.7434 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,380B, BPFP=2.6912 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,324B, BPFP=3.6838 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,660B, BPFP=2.0495 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,324B, BPFP=3.8501 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,576B, BPFP=2.2352 -⌛️ [2/4] FRONTEND: Frontend time: 1.809s (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: 1.377s - -[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, 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 0.03150173 4.43861844 - layer.0.v_cache 0.00000027 0.00006303 - layer.1.k_cache 0.00348105 0.36696945 - layer.1.v_cache 0.00000083 0.00019792 - layer.2.k_cache 0.00114206 0.17169133 - layer.2.v_cache 0.00000103 0.00026026 - layer.3.k_cache 0.00133122 0.19719053 - layer.3.v_cache 0.00000211 0.00045491 - layer.4.k_cache 0.00332532 0.40086884 - layer.4.v_cache 0.00000306 0.00072386 - layer.4.output 0.00018383 0.05298275 - ------------------------------------------------------------------------------------- - TOTAL 0.00296600 0.41349782 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 457860 -BPFP 2.7181 bits/point -EBPFP 5.4362 equivalent bits/point -MSE 0.413498 ----------------------- -------------------------------------------------------- -Time: 3.193s Load: 0.007s, Pack+Encode: 1.809s, Decode+Unpack: 1.377s ----------------------- -------------------------------------------------------- -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 0.4135 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample67-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample67-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample68-layer4-item1.zst (77/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample68-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 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, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,984B, BPFP=1.1147 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,560B, BPFP=3.4821 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,036B, BPFP=2.0635 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,524B, BPFP=3.7366 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,920B, BPFP=2.5687 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,816B, BPFP=3.7617 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,576B, BPFP=2.7109 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,712B, BPFP=3.6669 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,600B, BPFP=2.0261 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,976B, BPFP=3.8613 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,760B, BPFP=2.2914 -⌛️ [2/4] FRONTEND: Frontend time: 1.766s (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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.389s - -[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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03172423 4.39504301 - layer.0.v_cache 0.00000027 0.00006551 - layer.1.k_cache 0.00338285 0.32184517 - layer.1.v_cache 0.00000085 0.00020020 - layer.2.k_cache 0.00114373 0.16823075 - layer.2.v_cache 0.00000106 0.00027221 - layer.3.k_cache 0.00134976 0.19436163 - layer.3.v_cache 0.00000205 0.00046882 - layer.4.k_cache 0.00321980 0.40295481 - layer.4.v_cache 0.00000306 0.00077175 - layer.4.output 0.00016531 0.03717987 - ------------------------------------------------------------------------------------- - TOTAL 0.00296349 0.40235238 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 444464 -BPFP 2.7256 bits/point -EBPFP 5.4511 equivalent bits/point -MSE 0.402352 ----------------------- -------------------------------------------------------- -Time: 3.162s Load: 0.006s, Pack+Encode: 1.766s, Decode+Unpack: 1.389s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4024 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample68-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample68-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample7-layer4-item1.zst (78/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample7-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: 0.007s - ------------------------------------------------------------- -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: 12,812B, BPFP=1.0763 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,480B, BPFP=3.4845 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,464B, BPFP=2.1391 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,340B, BPFP=3.8088 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,660B, BPFP=2.6596 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,952B, BPFP=3.7762 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,244B, BPFP=2.7087 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,808B, BPFP=3.6801 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,592B, BPFP=2.0659 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,972B, BPFP=3.8619 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,672B, BPFP=2.2403 -⌛️ [2/4] FRONTEND: Frontend time: 1.816s (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: 1.342s - -[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 0.03162601 4.76223066 - layer.0.v_cache 0.00000026 0.00006414 - layer.1.k_cache 0.00335281 0.35571913 - layer.1.v_cache 0.00000078 0.00019232 - layer.2.k_cache 0.00111993 0.16991959 - layer.2.v_cache 0.00000106 0.00027001 - layer.3.k_cache 0.00131014 0.19514519 - layer.3.v_cache 0.00000197 0.00044342 - layer.4.k_cache 0.00327871 0.39595188 - layer.4.v_cache 0.00000299 0.00075493 - layer.4.output 0.00016230 0.05056200 - ------------------------------------------------------------------------------------- - TOTAL 0.00295313 0.43449566 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 454996 -BPFP 2.7302 bits/point -EBPFP 5.4603 equivalent bits/point -MSE 0.434496 ----------------------- -------------------------------------------------------- -Time: 3.166s Load: 0.007s, Pack+Encode: 1.816s, Decode+Unpack: 1.342s ----------------------- -------------------------------------------------------- -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 0.4345 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample7-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample7-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample70-layer4-item1.zst (79/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample70-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: 13,292B, BPFP=1.1047 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,168B, BPFP=3.5047 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,784B, BPFP=2.2261 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,932B, BPFP=3.7344 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,044B, BPFP=2.6632 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,832B, BPFP=3.7261 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,816B, BPFP=2.7274 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,140B, BPFP=3.6686 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,240B, BPFP=2.1809 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,440B, BPFP=3.7766 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 111,212B, BPFP=2.3108 -⌛️ [2/4] FRONTEND: Frontend time: 1.799s (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: 1.336s - -[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 0.03152096 4.17988814 - layer.0.v_cache 0.00000028 0.00006432 - layer.1.k_cache 0.00374224 0.35001024 - layer.1.v_cache 0.00000083 0.00020946 - layer.2.k_cache 0.00112812 0.16927547 - layer.2.v_cache 0.00000115 0.00028879 - layer.3.k_cache 0.00131826 0.19437339 - layer.3.v_cache 0.00000212 0.00047440 - layer.4.k_cache 0.00331939 0.42702987 - layer.4.v_cache 0.00000319 0.00079477 - layer.4.output 0.00017462 0.04889940 - ------------------------------------------------------------------------------------- - TOTAL 0.00298107 0.39414332 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 463900 -BPFP 2.7540 bits/point -EBPFP 5.5079 equivalent bits/point -MSE 0.394143 ----------------------- -------------------------------------------------------- -Time: 3.141s Load: 0.006s, Pack+Encode: 1.799s, Decode+Unpack: 1.336s ----------------------- -------------------------------------------------------- -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 0.3941 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample70-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample70-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample71-layer4-item1.zst (80/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample71-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: 13,128B, BPFP=1.1148 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,660B, BPFP=3.5377 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,028B, BPFP=2.1253 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,152B, BPFP=3.7493 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,156B, BPFP=2.6457 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,460B, BPFP=3.8604 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,648B, BPFP=2.7724 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,412B, BPFP=3.7714 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,764B, BPFP=2.1029 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,916B, BPFP=3.8991 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,012B, BPFP=2.2931 -⌛️ [2/4] FRONTEND: Frontend time: 1.822s (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: 1.341s - -[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 0.03010495 4.91113215 - layer.0.v_cache 0.00000027 0.00006741 - layer.1.k_cache 0.00345006 0.33927439 - layer.1.v_cache 0.00000091 0.00020786 - layer.2.k_cache 0.00113922 0.17027625 - layer.2.v_cache 0.00000110 0.00028102 - layer.3.k_cache 0.00129336 0.19944664 - layer.3.v_cache 0.00000204 0.00047654 - layer.4.k_cache 0.00332675 0.40754210 - layer.4.v_cache 0.00000304 0.00075663 - layer.4.output 0.00015992 0.03520726 - ------------------------------------------------------------------------------------- - TOTAL 0.00285438 0.44073500 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 456336 -BPFP 2.7680 bits/point -EBPFP 5.5359 equivalent bits/point -MSE 0.440735 ----------------------- -------------------------------------------------------- -Time: 3.169s Load: 0.006s, Pack+Encode: 1.822s, Decode+Unpack: 1.341s ----------------------- -------------------------------------------------------- -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 0.4407 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample71-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample71-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample72-layer4-item1.zst (81/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample72-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: 0.007s - ------------------------------------------------------------- -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: 12,932B, BPFP=1.0864 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,400B, BPFP=3.4778 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,092B, BPFP=2.1079 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,516B, BPFP=3.8236 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,636B, BPFP=2.6576 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,364B, BPFP=3.8108 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,348B, BPFP=2.6334 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,352B, BPFP=3.7258 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,188B, BPFP=2.1159 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,652B, BPFP=3.8350 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,172B, BPFP=2.2508 -⌛️ [2/4] FRONTEND: Frontend time: 1.789s (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: 1.351s - -[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 0.03144996 4.18264311 - layer.0.v_cache 0.00000027 0.00006272 - layer.1.k_cache 0.00343081 0.36163457 - layer.1.v_cache 0.00000081 0.00020277 - layer.2.k_cache 0.00115410 0.16797820 - layer.2.v_cache 0.00000107 0.00027184 - layer.3.k_cache 0.00131217 0.19459948 - layer.3.v_cache 0.00000205 0.00044689 - layer.4.k_cache 0.00325876 0.42223904 - layer.4.v_cache 0.00000310 0.00075907 - layer.4.output 0.00018378 0.05127347 - ------------------------------------------------------------------------------------- - TOTAL 0.00295345 0.39542369 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 455652 -BPFP 2.7341 bits/point -EBPFP 5.4682 equivalent bits/point -MSE 0.395424 ----------------------- -------------------------------------------------------- -Time: 3.147s Load: 0.007s, Pack+Encode: 1.789s, Decode+Unpack: 1.351s ----------------------- -------------------------------------------------------- -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 0.3954 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample72-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample72-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample73-layer4-item1.zst (82/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample73-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: 12,692B, BPFP=1.1665 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 39,668B, BPFP=3.6460 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,896B, BPFP=2.1044 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,644B, BPFP=3.8276 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,304B, BPFP=2.6015 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,572B, BPFP=3.9129 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,860B, BPFP=2.6526 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,604B, BPFP=3.8239 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,276B, BPFP=2.0474 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 41,960B, BPFP=3.8566 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,424B, BPFP=2.3994 -⌛️ [2/4] FRONTEND: Frontend time: 1.776s (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: 1.335s - -[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 0.03191245 4.28917847 - layer.0.v_cache 0.00000028 0.00006233 - layer.1.k_cache 0.00355479 0.36312837 - layer.1.v_cache 0.00000082 0.00021240 - layer.2.k_cache 0.00115713 0.16621915 - layer.2.v_cache 0.00000109 0.00028359 - layer.3.k_cache 0.00128442 0.18605028 - layer.3.v_cache 0.00000206 0.00048301 - layer.4.k_cache 0.00320041 0.38247717 - layer.4.v_cache 0.00000307 0.00078133 - layer.4.output 0.00016954 0.04842200 - ------------------------------------------------------------------------------------- - TOTAL 0.00298533 0.39875458 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 426900 -BPFP 2.8027 bits/point -EBPFP 5.6053 equivalent bits/point -MSE 0.398755 ----------------------- -------------------------------------------------------- -Time: 3.118s Load: 0.006s, Pack+Encode: 1.776s, Decode+Unpack: 1.335s ----------------------- -------------------------------------------------------- -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 0.3988 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample73-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample73-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample74-layer4-item1.zst (83/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample74-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: 0.006s - ------------------------------------------------------------- -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: 13,136B, BPFP=1.1403 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,100B, BPFP=3.5677 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,088B, BPFP=2.0910 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,556B, BPFP=3.8677 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,180B, BPFP=2.6198 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,912B, BPFP=3.8986 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,236B, BPFP=2.7983 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,872B, BPFP=3.7215 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,768B, BPFP=2.1500 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,396B, BPFP=3.9406 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 109,540B, BPFP=2.3772 -⌛️ [2/4] FRONTEND: Frontend time: 1.802s (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: 1.367s - -[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 0.03149115 4.63504605 - layer.0.v_cache 0.00000027 0.00006312 - layer.1.k_cache 0.00351934 0.34481591 - layer.1.v_cache 0.00000089 0.00020135 - layer.2.k_cache 0.00114079 0.16814122 - layer.2.v_cache 0.00000110 0.00027982 - layer.3.k_cache 0.00129789 0.19795960 - layer.3.v_cache 0.00000226 0.00048981 - layer.4.k_cache 0.00328524 0.41148071 - layer.4.v_cache 0.00000318 0.00078175 - layer.4.output 0.00020520 0.04574647 - ------------------------------------------------------------------------------------- - TOTAL 0.00296878 0.42444609 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 452784 -BPFP 2.8074 bits/point -EBPFP 5.6149 equivalent bits/point -MSE 0.424446 ----------------------- -------------------------------------------------------- -Time: 3.176s Load: 0.006s, Pack+Encode: 1.802s, Decode+Unpack: 1.367s ----------------------- -------------------------------------------------------- -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 0.4244 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample74-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample74-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample75-layer4-item1.zst (84/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample75-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: 13,044B, BPFP=1.1077 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,420B, BPFP=3.5173 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,668B, BPFP=2.1797 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,800B, BPFP=3.8043 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,312B, BPFP=2.6590 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,588B, BPFP=3.7863 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,568B, BPFP=2.7656 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,668B, BPFP=3.7931 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,744B, BPFP=2.1012 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,852B, BPFP=3.8937 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,560B, BPFP=2.2835 -⌛️ [2/4] FRONTEND: Frontend time: 1.785s (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: 1.326s - -[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 0.03262251 4.63584767 - layer.0.v_cache 0.00000028 0.00006362 - layer.1.k_cache 0.00349940 0.35791347 - layer.1.v_cache 0.00000082 0.00020261 - layer.2.k_cache 0.00113634 0.17090401 - layer.2.v_cache 0.00000107 0.00027705 - layer.3.k_cache 0.00132950 0.19262046 - layer.3.v_cache 0.00000207 0.00048278 - layer.4.k_cache 0.00323306 0.40982251 - layer.4.v_cache 0.00000312 0.00078414 - layer.4.output 0.00016774 0.04640522 - ------------------------------------------------------------------------------------- - TOTAL 0.00303565 0.42532423 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 456224 -BPFP 2.7673 bits/point -EBPFP 5.5345 equivalent bits/point -MSE 0.425324 ----------------------- -------------------------------------------------------- -Time: 3.117s Load: 0.006s, Pack+Encode: 1.785s, Decode+Unpack: 1.326s ----------------------- -------------------------------------------------------- -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 0.4253 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample75-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample75-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample78-layer4-item1.zst (85/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample78-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 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, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,928B, BPFP=1.1099 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,376B, BPFP=3.5522 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,344B, BPFP=2.0900 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,312B, BPFP=3.7184 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,680B, BPFP=2.6339 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,712B, BPFP=3.8386 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,328B, BPFP=2.7754 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,948B, BPFP=3.6872 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,068B, BPFP=2.0663 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,564B, BPFP=3.9117 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,844B, BPFP=2.2717 -⌛️ [2/4] FRONTEND: Frontend time: 1.790s (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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.343s - -[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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03087558 4.53481687 - layer.0.v_cache 0.00000028 0.00006417 - layer.1.k_cache 0.00335485 0.31529464 - layer.1.v_cache 0.00000080 0.00019638 - layer.2.k_cache 0.00115595 0.16810141 - layer.2.v_cache 0.00000105 0.00026732 - layer.3.k_cache 0.00133087 0.19918622 - layer.3.v_cache 0.00000204 0.00045062 - layer.4.k_cache 0.00324117 0.40233838 - layer.4.v_cache 0.00000319 0.00077757 - layer.4.output 0.00016760 0.04076485 - ------------------------------------------------------------------------------------- - TOTAL 0.00290258 0.41318236 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 448104 -BPFP 2.7479 bits/point -EBPFP 5.4958 equivalent bits/point -MSE 0.413182 ----------------------- -------------------------------------------------------- -Time: 3.139s Load: 0.006s, Pack+Encode: 1.790s, Decode+Unpack: 1.343s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4132 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample78-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample78-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample8-layer4-item1.zst (86/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample8-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: 13,292B, BPFP=1.0706 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 43,748B, BPFP=3.5235 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,712B, BPFP=2.1514 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,840B, BPFP=3.6920 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,104B, BPFP=2.6662 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,960B, BPFP=3.7017 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,536B, BPFP=2.7010 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,588B, BPFP=3.6717 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,440B, BPFP=2.1295 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 47,488B, BPFP=3.8247 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 114,984B, BPFP=2.3152 -⌛️ [2/4] FRONTEND: Frontend time: 1.775s (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: 1.350s - -[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 0.03000178 4.58257931 - layer.0.v_cache 0.00000027 0.00006249 - layer.1.k_cache 0.00327340 0.35025662 - layer.1.v_cache 0.00000092 0.00020047 - layer.2.k_cache 0.00113100 0.17546093 - layer.2.v_cache 0.00000104 0.00026799 - layer.3.k_cache 0.00131735 0.19746360 - layer.3.v_cache 0.00000202 0.00044485 - layer.4.k_cache 0.00333574 0.43255324 - layer.4.v_cache 0.00000307 0.00076189 - layer.4.output 0.00017322 0.04702985 - ------------------------------------------------------------------------------------- - TOTAL 0.00283996 0.42344077 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 476692 -BPFP 2.7424 bits/point -EBPFP 5.4848 equivalent bits/point -MSE 0.423441 ----------------------- -------------------------------------------------------- -Time: 3.133s Load: 0.007s, Pack+Encode: 1.775s, Decode+Unpack: 1.350s ----------------------- -------------------------------------------------------- -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 0.4234 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample8-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample8-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample80-layer4-item1.zst (87/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample80-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 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, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,848B, BPFP=1.1030 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,932B, BPFP=3.5141 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,944B, BPFP=2.0556 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,536B, BPFP=3.7376 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,196B, BPFP=2.5924 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,068B, BPFP=3.7833 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,968B, BPFP=2.7445 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,952B, BPFP=3.6875 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,348B, BPFP=2.0045 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,196B, BPFP=3.8802 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,648B, BPFP=2.2461 -⌛️ [2/4] FRONTEND: Frontend time: 1.759s (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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.351s - -[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, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03041337 4.54467103 - layer.0.v_cache 0.00000027 0.00006524 - layer.1.k_cache 0.00341539 0.33338211 - layer.1.v_cache 0.00000079 0.00019902 - layer.2.k_cache 0.00115498 0.16927061 - layer.2.v_cache 0.00000106 0.00028147 - layer.3.k_cache 0.00132454 0.19338140 - layer.3.v_cache 0.00000206 0.00045112 - layer.4.k_cache 0.00329878 0.40072175 - layer.4.v_cache 0.00000306 0.00074830 - layer.4.output 0.00018052 0.03904180 - ------------------------------------------------------------------------------------- - TOTAL 0.00288117 0.41423852 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 443636 -BPFP 2.7205 bits/point -EBPFP 5.4410 equivalent bits/point -MSE 0.414239 ----------------------- -------------------------------------------------------- -Time: 3.117s Load: 0.007s, Pack+Encode: 1.759s, Decode+Unpack: 1.351s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4142 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample80-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample80-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample81-layer4-item1.zst (88/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample81-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 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, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 12,560B, BPFP=1.1151 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,448B, BPFP=3.5909 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,640B, BPFP=2.0987 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,552B, BPFP=3.7777 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,612B, BPFP=2.5401 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,228B, BPFP=3.8377 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,668B, BPFP=2.7227 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,596B, BPFP=3.6928 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,124B, BPFP=2.0529 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,060B, BPFP=3.8228 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,892B, BPFP=2.3502 -⌛️ [2/4] FRONTEND: Frontend time: 1.827s (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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.354s - -[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, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03108732 4.67954462 - layer.0.v_cache 0.00000028 0.00006397 - layer.1.k_cache 0.00347999 0.33093491 - layer.1.v_cache 0.00000081 0.00020605 - layer.2.k_cache 0.00113734 0.16505730 - layer.2.v_cache 0.00000108 0.00028232 - layer.3.k_cache 0.00129639 0.19746845 - layer.3.v_cache 0.00000204 0.00047187 - layer.4.k_cache 0.00326982 0.39033595 - layer.4.v_cache 0.00000302 0.00077567 - layer.4.output 0.00017394 0.04287664 - ------------------------------------------------------------------------------------- - TOTAL 0.00292670 0.42404626 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 435380 -BPFP 2.7609 bits/point -EBPFP 5.5218 equivalent bits/point -MSE 0.424046 ----------------------- -------------------------------------------------------- -Time: 3.188s Load: 0.007s, Pack+Encode: 1.827s, Decode+Unpack: 1.354s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.4240 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample81-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample81-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample82-layer4-item1.zst (89/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample82-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: 0.007s - ------------------------------------------------------------- -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: 13,092B, BPFP=1.1118 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,432B, BPFP=3.5183 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,044B, BPFP=2.1267 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,728B, BPFP=3.7982 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,924B, BPFP=2.6260 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,120B, BPFP=3.8315 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,420B, BPFP=2.7531 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,120B, BPFP=3.7466 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,996B, BPFP=2.1226 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,436B, BPFP=3.9433 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,592B, BPFP=2.3054 -⌛️ [2/4] FRONTEND: Frontend time: 1.774s (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: 1.344s - -[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 0.03045331 4.30085821 - layer.0.v_cache 0.00000027 0.00006193 - layer.1.k_cache 0.00344094 0.33570939 - layer.1.v_cache 0.00000081 0.00019979 - layer.2.k_cache 0.00113827 0.16971803 - layer.2.v_cache 0.00000106 0.00026973 - layer.3.k_cache 0.00131047 0.19176740 - layer.3.v_cache 0.00000202 0.00046083 - layer.4.k_cache 0.00338451 0.40413765 - layer.4.v_cache 0.00000307 0.00075006 - layer.4.output 0.00016907 0.04073902 - ------------------------------------------------------------------------------------- - TOTAL 0.00288650 0.39763494 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 456904 -BPFP 2.7714 bits/point -EBPFP 5.5428 equivalent bits/point -MSE 0.397635 ----------------------- -------------------------------------------------------- -Time: 3.125s Load: 0.007s, Pack+Encode: 1.774s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -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 0.3976 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample82-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample82-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample83-layer4-item1.zst (90/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample83-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: 0.007s - ------------------------------------------------------------- -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: 13,072B, BPFP=1.1101 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,500B, BPFP=3.5241 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,996B, BPFP=2.1226 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,748B, BPFP=3.7999 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,048B, BPFP=2.6365 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,108B, BPFP=3.8305 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,824B, BPFP=2.7874 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,112B, BPFP=3.7459 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,700B, BPFP=2.0975 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,264B, BPFP=3.9287 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,232B, BPFP=2.2977 -⌛️ [2/4] FRONTEND: Frontend time: 1.801s (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: 1.351s - -[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 0.03019064 4.28411666 - layer.0.v_cache 0.00000027 0.00006239 - layer.1.k_cache 0.00347154 0.35007419 - layer.1.v_cache 0.00000080 0.00019893 - layer.2.k_cache 0.00112864 0.16907038 - layer.2.v_cache 0.00000106 0.00027219 - layer.3.k_cache 0.00131097 0.19221764 - layer.3.v_cache 0.00000203 0.00045655 - layer.4.k_cache 0.00338659 0.40330460 - layer.4.v_cache 0.00000308 0.00074296 - layer.4.output 0.00016864 0.04033224 - ------------------------------------------------------------------------------------- - TOTAL 0.00286930 0.39727468 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 456604 -BPFP 2.7696 bits/point -EBPFP 5.5392 equivalent bits/point -MSE 0.397275 ----------------------- -------------------------------------------------------- -Time: 3.158s Load: 0.007s, Pack+Encode: 1.801s, Decode+Unpack: 1.351s ----------------------- -------------------------------------------------------- -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 0.3973 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample83-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample83-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample85-layer4-item1.zst (91/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample85-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: 13,256B, BPFP=1.0788 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 42,936B, BPFP=3.4941 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 26,548B, BPFP=2.1605 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,232B, BPFP=3.6810 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,740B, BPFP=2.6644 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,224B, BPFP=3.7617 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 33,376B, BPFP=2.7161 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,680B, BPFP=3.6361 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,956B, BPFP=2.1937 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,532B, BPFP=3.7868 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 111,116B, BPFP=2.2607 -⌛️ [2/4] FRONTEND: Frontend time: 1.796s (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: 1.357s - -[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 0.03109797 4.51961454 - layer.0.v_cache 0.00000027 0.00006396 - layer.1.k_cache 0.00337694 0.35192768 - layer.1.v_cache 0.00000077 0.00019437 - layer.2.k_cache 0.00112303 0.17024720 - layer.2.v_cache 0.00000110 0.00028137 - layer.3.k_cache 0.00132768 0.19637080 - layer.3.v_cache 0.00000201 0.00045754 - layer.4.k_cache 0.00329028 0.45297341 - layer.4.v_cache 0.00000306 0.00075735 - layer.4.output 0.00016342 0.04057163 - ------------------------------------------------------------------------------------- - TOTAL 0.00291977 0.41822677 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 469596 -BPFP 2.7297 bits/point -EBPFP 5.4594 equivalent bits/point -MSE 0.418227 ----------------------- -------------------------------------------------------- -Time: 3.159s Load: 0.006s, Pack+Encode: 1.796s, Decode+Unpack: 1.357s ----------------------- -------------------------------------------------------- -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 0.4182 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample85-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample85-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample88-layer4-item1.zst (92/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample88-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: 0.007s - ------------------------------------------------------------- -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: 13,112B, BPFP=1.1015 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,500B, BPFP=3.4862 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,544B, BPFP=2.1458 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 45,100B, BPFP=3.7886 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,480B, BPFP=2.6445 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,652B, BPFP=3.8350 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,908B, BPFP=2.6804 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,816B, BPFP=3.6808 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,292B, BPFP=2.0407 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,216B, BPFP=3.7984 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,668B, BPFP=2.2402 -⌛️ [2/4] FRONTEND: Frontend time: 1.812s (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: 1.330s - -[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 0.03084236 4.38202987 - layer.0.v_cache 0.00000028 0.00006397 - layer.1.k_cache 0.00343721 0.36558479 - layer.1.v_cache 0.00000078 0.00019934 - layer.2.k_cache 0.00115435 0.17209088 - layer.2.v_cache 0.00000107 0.00027730 - layer.3.k_cache 0.00133139 0.20111597 - layer.3.v_cache 0.00000204 0.00047038 - layer.4.k_cache 0.00330969 0.40692258 - layer.4.v_cache 0.00000309 0.00077197 - layer.4.output 0.00020550 0.05364104 - ------------------------------------------------------------------------------------- - TOTAL 0.00292173 0.41029223 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 454288 -BPFP 2.7259 bits/point -EBPFP 5.4518 equivalent bits/point -MSE 0.410292 ----------------------- -------------------------------------------------------- -Time: 3.148s Load: 0.007s, Pack+Encode: 1.812s, Decode+Unpack: 1.330s ----------------------- -------------------------------------------------------- -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 0.4103 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample88-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample88-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample90-layer4-item1.zst (93/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample90-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: 0.008s - ------------------------------------------------------------- -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: 12,808B, BPFP=1.1243 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,904B, BPFP=3.5906 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,824B, BPFP=2.0913 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,712B, BPFP=3.8371 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,192B, BPFP=2.5625 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,352B, BPFP=3.8055 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,424B, BPFP=2.7584 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,176B, BPFP=3.7022 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,976B, BPFP=2.0169 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,536B, BPFP=3.8216 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,988B, BPFP=2.3479 -⌛️ [2/4] FRONTEND: Frontend time: 1.861s (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: 1.334s - -[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 0.03125147 4.47739196 - layer.0.v_cache 0.00000028 0.00006115 - layer.1.k_cache 0.00348025 0.34083763 - layer.1.v_cache 0.00000083 0.00020163 - layer.2.k_cache 0.00111536 0.16883404 - layer.2.v_cache 0.00000108 0.00027349 - layer.3.k_cache 0.00131582 0.19718732 - layer.3.v_cache 0.00000206 0.00045370 - layer.4.k_cache 0.00329313 0.40794154 - layer.4.v_cache 0.00000300 0.00074582 - layer.4.output 0.00017131 0.04728290 - ------------------------------------------------------------------------------------- - TOTAL 0.00293918 0.41307571 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 440892 -BPFP 2.7644 bits/point -EBPFP 5.5288 equivalent bits/point -MSE 0.413076 ----------------------- -------------------------------------------------------- -Time: 3.202s Load: 0.008s, Pack+Encode: 1.861s, Decode+Unpack: 1.334s ----------------------- -------------------------------------------------------- -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 0.4131 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample90-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample90-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample92-layer4-item1.zst (94/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample92-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: 13,032B, BPFP=1.0948 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,564B, BPFP=3.4916 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,460B, BPFP=2.1388 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,936B, BPFP=3.7749 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,472B, BPFP=2.6438 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 45,160B, BPFP=3.7937 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 31,488B, BPFP=2.6452 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,180B, BPFP=3.7114 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,776B, BPFP=2.0813 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,684B, BPFP=3.8377 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 108,684B, BPFP=2.2825 -⌛️ [2/4] FRONTEND: Frontend time: 1.821s (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: 1.370s - -[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 0.03101371 4.13054009 - layer.0.v_cache 0.00000028 0.00006072 - layer.1.k_cache 0.00364020 0.35344450 - layer.1.v_cache 0.00000086 0.00019802 - layer.2.k_cache 0.00114853 0.17299408 - layer.2.v_cache 0.00000108 0.00026850 - layer.3.k_cache 0.00131058 0.19021838 - layer.3.v_cache 0.00000204 0.00044262 - layer.4.k_cache 0.00330942 0.42178853 - layer.4.v_cache 0.00000298 0.00076614 - layer.4.output 0.00017313 0.04854604 - ------------------------------------------------------------------------------------- - TOTAL 0.00293730 0.39035041 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 456436 -BPFP 2.7388 bits/point -EBPFP 5.4776 equivalent bits/point -MSE 0.390350 ----------------------- -------------------------------------------------------- -Time: 3.198s Load: 0.006s, Pack+Encode: 1.821s, Decode+Unpack: 1.370s ----------------------- -------------------------------------------------------- -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 0.3904 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample92-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample92-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample93-layer4-item1.zst (95/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample93-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: 12,692B, BPFP=1.1017 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,240B, BPFP=3.5799 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,068B, BPFP=2.0892 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,320B, BPFP=3.7604 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,288B, BPFP=2.5424 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,948B, BPFP=3.8149 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,008B, BPFP=2.7785 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,040B, BPFP=3.7361 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,936B, BPFP=2.0778 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,780B, BPFP=3.9740 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 103,860B, BPFP=2.2539 -⌛️ [2/4] FRONTEND: Frontend time: 1.740s (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: 1.335s - -[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 0.03189978 4.63139174 - layer.0.v_cache 0.00000027 0.00006483 - layer.1.k_cache 0.00346617 0.35751093 - layer.1.v_cache 0.00000078 0.00019753 - layer.2.k_cache 0.00115535 0.16727319 - layer.2.v_cache 0.00000103 0.00026598 - layer.3.k_cache 0.00134436 0.20336825 - layer.3.v_cache 0.00000198 0.00044009 - layer.4.k_cache 0.00332606 0.43147210 - layer.4.v_cache 0.00000301 0.00075901 - layer.4.output 0.00018488 0.05517720 - ------------------------------------------------------------------------------------- - TOTAL 0.00299559 0.42953232 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 443180 -BPFP 2.7479 bits/point -EBPFP 5.4958 equivalent bits/point -MSE 0.429532 ----------------------- -------------------------------------------------------- -Time: 3.080s Load: 0.005s, Pack+Encode: 1.740s, Decode+Unpack: 1.335s ----------------------- -------------------------------------------------------- -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 0.4295 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample93-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample93-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample94-layer4-item1.zst (96/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample94-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: 12,624B, BPFP=1.1081 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,396B, BPFP=3.5460 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,920B, BPFP=2.0997 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,272B, BPFP=3.7985 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,408B, BPFP=2.5815 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,120B, BPFP=3.7851 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,580B, BPFP=2.6843 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,112B, BPFP=3.6966 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,176B, BPFP=2.0344 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,580B, BPFP=3.8255 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,620B, BPFP=2.3179 -⌛️ [2/4] FRONTEND: Frontend time: 1.766s (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: 1.329s - -[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 0.03210033 4.51630805 - layer.0.v_cache 0.00000027 0.00006221 - layer.1.k_cache 0.00338447 0.34458315 - layer.1.v_cache 0.00000081 0.00020607 - layer.2.k_cache 0.00116617 0.16656297 - layer.2.v_cache 0.00000108 0.00028456 - layer.3.k_cache 0.00134809 0.19386583 - layer.3.v_cache 0.00000208 0.00048124 - layer.4.k_cache 0.00324151 0.39948496 - layer.4.v_cache 0.00000296 0.00075590 - layer.4.output 0.00019744 0.04620977 - ------------------------------------------------------------------------------------- - TOTAL 0.00300268 0.41481672 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 437808 -BPFP 2.7451 bits/point -EBPFP 5.4902 equivalent bits/point -MSE 0.414817 ----------------------- -------------------------------------------------------- -Time: 3.101s Load: 0.006s, Pack+Encode: 1.766s, Decode+Unpack: 1.329s ----------------------- -------------------------------------------------------- -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 0.4148 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample94-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample94-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample95-layer4-item1.zst (97/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample95-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: 12,684B, BPFP=1.1134 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,552B, BPFP=3.5597 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,668B, BPFP=2.0776 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,352B, BPFP=3.8055 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,148B, BPFP=2.5586 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,236B, BPFP=3.7953 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,684B, BPFP=2.6935 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,124B, BPFP=3.6977 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,336B, BPFP=2.0485 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,684B, BPFP=3.8346 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 105,736B, BPFP=2.3204 -⌛️ [2/4] FRONTEND: Frontend time: 1.761s (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: 1.337s - -[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 0.03204785 4.50029660 - layer.0.v_cache 0.00000027 0.00006227 - layer.1.k_cache 0.00337303 0.33991162 - layer.1.v_cache 0.00000080 0.00020559 - layer.2.k_cache 0.00115601 0.17075579 - layer.2.v_cache 0.00000108 0.00028361 - layer.3.k_cache 0.00135783 0.19744402 - layer.3.v_cache 0.00000209 0.00048246 - layer.4.k_cache 0.00324176 0.39556225 - layer.4.v_cache 0.00000298 0.00075618 - layer.4.output 0.00019437 0.04866541 - ------------------------------------------------------------------------------------- - TOTAL 0.00299723 0.41431586 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 438204 -BPFP 2.7476 bits/point -EBPFP 5.4951 equivalent bits/point -MSE 0.414316 ----------------------- -------------------------------------------------------- -Time: 3.105s Load: 0.007s, Pack+Encode: 1.761s, Decode+Unpack: 1.337s ----------------------- -------------------------------------------------------- -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 0.4143 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample95-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample95-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample96-layer4-item1.zst (98/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample96-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: 12,540B, BPFP=1.1526 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 39,548B, BPFP=3.6349 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 22,948B, BPFP=2.1092 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,748B, BPFP=3.8371 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,888B, BPFP=2.4713 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,636B, BPFP=3.9188 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,848B, BPFP=2.6515 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,848B, BPFP=3.7544 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,224B, BPFP=2.0426 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,836B, BPFP=3.9371 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 102,600B, BPFP=2.3575 -⌛️ [2/4] FRONTEND: Frontend time: 1.792s (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: 1.362s - -[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 0.03332241 4.65950641 - layer.0.v_cache 0.00000028 0.00006510 - layer.1.k_cache 0.00345674 0.34989700 - layer.1.v_cache 0.00000080 0.00020832 - layer.2.k_cache 0.00115406 0.16299986 - layer.2.v_cache 0.00000108 0.00028360 - layer.3.k_cache 0.00129640 0.18277617 - layer.3.v_cache 0.00000208 0.00047016 - layer.4.k_cache 0.00325461 0.38520858 - layer.4.v_cache 0.00000299 0.00077196 - layer.4.output 0.00016656 0.04651256 - ------------------------------------------------------------------------------------- - TOTAL 0.00308269 0.42344553 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 423664 -BPFP 2.7814 bits/point -EBPFP 5.5628 equivalent bits/point -MSE 0.423446 ----------------------- -------------------------------------------------------- -Time: 3.161s Load: 0.006s, Pack+Encode: 1.792s, Decode+Unpack: 1.362s ----------------------- -------------------------------------------------------- -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 0.4234 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample96-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample96-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample98-layer4-item1.zst (99/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample98-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: 13,028B, BPFP=1.1063 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 41,588B, BPFP=3.5316 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,352B, BPFP=2.1529 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,636B, BPFP=3.7904 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 31,024B, BPFP=2.6345 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,980B, BPFP=3.8196 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 32,604B, BPFP=2.7687 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 44,640B, BPFP=3.7908 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,504B, BPFP=2.1658 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,148B, BPFP=3.9188 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 107,756B, BPFP=2.2876 -⌛️ [2/4] FRONTEND: Frontend time: 1.741s (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: 1.339s - -[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 0.03086128 4.44609766 - layer.0.v_cache 0.00000027 0.00006428 - layer.1.k_cache 0.00344194 0.32758570 - layer.1.v_cache 0.00000079 0.00019426 - layer.2.k_cache 0.00114743 0.16737432 - layer.2.v_cache 0.00000107 0.00027267 - layer.3.k_cache 0.00130642 0.19277653 - layer.3.v_cache 0.00000205 0.00044920 - layer.4.k_cache 0.00331827 0.42506201 - layer.4.v_cache 0.00000300 0.00073406 - layer.4.output 0.00017802 0.04519708 - ------------------------------------------------------------------------------------- - TOTAL 0.00291390 0.41009993 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 457260 -BPFP 2.7736 bits/point -EBPFP 5.5471 equivalent bits/point -MSE 0.410100 ----------------------- -------------------------------------------------------- -Time: 3.085s Load: 0.006s, Pack+Encode: 1.741s, Decode+Unpack: 1.339s ----------------------- -------------------------------------------------------- -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 0.4101 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample98-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample98-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample99-layer4-item1.zst (100/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample99-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: 12,868B, BPFP=1.1555 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 40,020B, BPFP=3.5938 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,428B, BPFP=2.1038 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,400B, BPFP=3.8075 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,416B, BPFP=2.5517 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,520B, BPFP=3.9080 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 30,252B, BPFP=2.7166 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 41,760B, BPFP=3.7500 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,684B, BPFP=2.0370 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,012B, BPFP=3.8624 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 106,604B, BPFP=2.3932 -⌛️ [2/4] FRONTEND: Frontend time: 1.764s (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: 1.383s - -[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 0.03104359 4.46920145 - layer.0.v_cache 0.00000028 0.00006366 - layer.1.k_cache 0.00347075 0.33117108 - layer.1.v_cache 0.00000085 0.00021013 - layer.2.k_cache 0.00119839 0.16577210 - layer.2.v_cache 0.00000112 0.00029173 - layer.3.k_cache 0.00133526 0.19091725 - layer.3.v_cache 0.00000211 0.00048350 - layer.4.k_cache 0.00323497 0.38733813 - layer.4.v_cache 0.00000315 0.00077699 - layer.4.output 0.00017081 0.03967880 - ------------------------------------------------------------------------------------- - TOTAL 0.00292669 0.40749580 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 434964 -BPFP 2.7899 bits/point -EBPFP 5.5799 equivalent bits/point -MSE 0.407496 ----------------------- -------------------------------------------------------- -Time: 3.153s Load: 0.006s, Pack+Encode: 1.764s, Decode+Unpack: 1.383s ----------------------- -------------------------------------------------------- -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 0.4075 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample99-layer4-item1.zst - to output-fixed/qwen/lambda0.02/elic-featurecoding-8bit-individual/fc_winogrande/sample99-layer4-item1.zst ------------------------- ---------------------------- -TOTAL PROCESSING SUMMARY ------------------------- ---------------------------- -Total files 100 -Avg BPFP 2.7560 bits/point -Avg EBPFP 5.5119 equivalent bits/point -Avg MSE 0.412069 -Avg Time 3.151s ------------------------- ---------------------------- +version https://git-lfs.github.com/spec/v1 +oid sha256:b982a4b74ff4ae0dd04dc1fb3c4618d0f82016b7981205bdbdf7d46349a161df +size 1117486