Upload SFT model: 26_09_08_sft_email_v4_v3

.gitattributes

@@ -56,3 +56,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 26_09_08_sft_email_v4/export/gemma4/120/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 26_09_08_sft_email_v4/export/gemma4/180/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 26_09_08_sft_email_v4/export/gemma4/60/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+26_09_08_sft_email_v4_v3/export/gemma4/140/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+26_09_08_sft_email_v4_v3/export/gemma4/210/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+26_09_08_sft_email_v4_v3/export/gemma4/70/tokenizer.json filter=lfs diff=lfs merge=lfs -text

26_09_08_sft_email_v4_v3/config.yaml

@@ -0,0 +1,267 @@
+data:
+  datasets:
+    training:
+      datasets:
+      - path: /mnt/queue1/shashank/exp/data/26_09_04_sft_email_v3/fastllm/fast_llm_config.yaml
+        type: file
+      - path: /mnt/queue1/shashank/exp/data/26_09_08_sft_email_v4/fastllm/fast_llm_config.yaml
+        type: file
+      - path: /mnt/queue1/shruthan/data/gemma_sft_replay/10k/fastllm/fast_llm_config.yaml
+        type: file
+      name: gemma_replay_and_automation_bench
+      type: blended
+      weights:
+      - 0.15008542238000852
+      - 0.3285671651047079
+      - 0.5213474125152836
+  maximum_document_length: 32768
+  micro_batch_size: 32768
+  truncate_documents: false
+model:
+  base_model:
+    decoder:
+      blocks:
+        full_attention:
+          mixer:
+            add_linear_biases: false
+            dropout: 0.0
+            head_groups: 2
+            head_size: 512
+            heads: 16
+            implementation: sdpa
+            key_norm:
+              epsilon: 1.0e-06
+              type: rms_norm
+            query_norm:
+              epsilon: 1.0e-06
+              type: rms_norm
+            rotary:
+              partial_rotary_factor: 0.25
+              theta: 1000000.0
+              type: proportional
+            shared_key_value: true
+            softmax_scale_power: 0.0
+            type: attention
+            value_norm:
+              epsilon: 1.0e-06
+              type: fixed_rms_norm
+          mlp:
+            dense:
+              activation: gelu
+              add_linear_biases: false
+              gated: true
+              intermediate_size: 2112
+              post_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              pre_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              recompute_level: full
+              type: mlp
+            routed:
+              activation: gelu
+              add_linear_biases: false
+              auxiliary_loss_coefficient: 0.0
+              experts: 128
+              experts_per_token: 8
+              gated: true
+              intermediate_size: 704
+              post_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              pre_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              recompute_level: full
+              router_input_scale: 0.018844459036110227
+              router_normalization:
+                epsilon: 1.0e-06
+                type: fixed_rms_norm
+              router_per_expert_scale:
+                enabled: true
+              router_scale:
+                enabled: true
+              type: moe
+            type: hybrid_moe
+          normalization:
+            epsilon: 1.0e-06
+            type: rms_norm
+          output_scale:
+            enabled: true
+            lr_scale: 0.0
+          post_mixer_normalization:
+            epsilon: 1.0e-06
+            type: rms_norm
+          post_mlp_normalization:
+            epsilon: 1.0e-06
+            type: rms_norm
+          pre_mlp_normalization:
+            type: none
+          type: decoder
+        sliding_attention:
+          mixer:
+            add_linear_biases: false
+            dropout: 0.0
+            head_groups: 8
+            head_size: 256
+            heads: 16
+            implementation: flash
+            key_norm:
+              epsilon: 1.0e-06
+              type: rms_norm
+            query_norm:
+              epsilon: 1.0e-06
+              type: rms_norm
+            rotary:
+              theta: 10000.0
+              type: default
+            softmax_scale_power: 0.0
+            type: attention
+            value_norm:
+              epsilon: 1.0e-06
+              type: fixed_rms_norm
+            window_size: 1024
+          mlp:
+            dense:
+              activation: gelu
+              add_linear_biases: false
+              gated: true
+              intermediate_size: 2112
+              post_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              pre_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              recompute_level: full
+              type: mlp
+            routed:
+              activation: gelu
+              add_linear_biases: false
+              auxiliary_loss_coefficient: 0.0
+              experts: 128
+              experts_per_token: 8
+              gated: true
+              intermediate_size: 704
+              post_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              pre_norm:
+                epsilon: 1.0e-06
+                type: rms_norm
+              recompute_level: full
+              router_input_scale: 0.018844459036110227
+              router_normalization:
+                epsilon: 1.0e-06
+                type: fixed_rms_norm
+              router_per_expert_scale:
+                enabled: true
+              router_scale:
+                enabled: true
+              type: moe
+            type: hybrid_moe
+          normalization:
+            epsilon: 1.0e-06
+            type: rms_norm
+          output_scale:
+            enabled: true
+            lr_scale: 0.0
+          post_mixer_normalization:
+            epsilon: 1.0e-06
+            type: rms_norm
+          post_mlp_normalization:
+            epsilon: 1.0e-06
+            type: rms_norm
+          pre_mlp_normalization:
+            type: none
+          type: decoder
+      num_blocks: 30
+      pattern:
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      type: pattern
+    embeddings:
+      embedding_scale: 53.0659966456864
+      vocab_parallel: false
+      vocab_size: 262144
+    head:
+      cross_entropy_splits: 8
+      final_logit_softcap: 30.0
+      normalization:
+        epsilon: 1.0e-06
+        type: rms_norm
+    hidden_size: 2816
+    peft:
+      type: none
+    tied_embedding_weight: true
+  distributed:
+    compute_dtype: bfloat16
+    local_world_size: 8
+    rank: 0
+    sequence_data_parallel: 4
+    sequence_tensor_parallel: true
+    tensor_parallel: 2
+    timeout: 3600.0
+    world_size: 32
+  multi_stage:
+    zero_stage: 3
+  type: gpt
+optimizer:
+  learning_rate:
+    base: 2.0e-05
+    decay_iterations: 210
+    decay_style: cosine
+    warmup_iterations: 7
+  weight_decay: 0.01
+pretrained:
+  format: gemma4
+  load_config: model
+  model_weights: true
+  path: /mnt/queue1/shruthan/models/gemma-4-26B-A4B-it-causal-lm
+run:
+  experiment_dir: /mnt/queue1/shashank/exp/sft/26_09_08_sft_email_v4_v3
+schedule:
+  breadth_first_micro_batches: 1
+  depth_first_micro_batches: 16
+training:
+  evaluators: {}
+  export:
+    format: gemma4
+    interval: 70
+  logs:
+    interval: 1
+  num_workers: 0
+  train_iters: 210
+  wandb:
+    entity_name: corellm
+    project_name: gemma-sft
+type: train_gpt

26_09_08_sft_email_v4_v3/config_verbose.yaml

@@ -0,0 +1,910 @@
+callbacks: {}
+data:
+  data_sample_warn_time_ms: 1000.0
+  datasets:
+    training:
+      datasets:
+      - path: /mnt/queue1/shashank/exp/data/26_09_04_sft_email_v3/fastllm/fast_llm_config.yaml
+        type: file
+      - path: /mnt/queue1/shashank/exp/data/26_09_08_sft_email_v4/fastllm/fast_llm_config.yaml
+        type: file
+      - path: /mnt/queue1/shruthan/data/gemma_sft_replay/10k/fastllm/fast_llm_config.yaml
+        type: file
+      name: gemma_replay_and_automation_bench
+      type: blended
+      weights:
+      - 0.15008542238000852
+      - 0.3285671651047079
+      - 0.5213474125152836
+  gpu: true
+  maximum_document_length: 32768
+  micro_batch_size: 32768
+  multiprocessing_context: spawn
+  seed: 784569
+  shuffle: epoch
+  truncate_documents: false
+  type: null
+model:
+  base_model:
+    decoder:
+      blocks:
+        full_attention:
+          distillation_loss_weight: 1.0
+          distillation_model: null
+          dropout: 0.0
+          lr_scale: null
+          mixer:
+            add_linear_biases: false
+            causal: true
+            dense_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            dropout: 0.0
+            head_groups: 2
+            head_size: 512
+            heads: 16
+            implementation: sdpa
+            key_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            key_norm:
+              epsilon: 1.0e-06
+              implementation: auto
+              lr_scale: null
+              type: rms_norm
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              zero_centered: false
+            lr_scale: null
+            query_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            query_norm:
+              epsilon: 1.0e-06
+              implementation: auto
+              lr_scale: null
+              type: rms_norm
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              zero_centered: false
+            rotary:
+              partial_rotary_factor: 0.25
+              theta: 1000000.0
+              type: proportional
+            shared_key_value: true
+            softmax_scale_power: 0.0
+            type: attention
+            use_flash_attention: true
+            value_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            value_norm:
+              epsilon: 1.0e-06
+              lr_scale: null
+              type: fixed_rms_norm
+            window_size: null
+          mlp:
+            dense:
+              activation: gelu
+              add_linear_biases: false
+              gated: true
+              intermediate_size: 2112
+              layer_1:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              layer_2:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              lr_scale: null
+              post_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              pre_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              recompute_level: full
+              type: mlp
+            lr_scale: null
+            post_norm: null
+            pre_norm: null
+            routed:
+              activation: gelu
+              add_linear_biases: false
+              auxiliary_loss_coefficient: 0.0
+              dropless_dynamic_shape: false
+              experts: 128
+              experts_per_token: 8
+              gated: true
+              implementation: auto
+              intermediate_size: 704
+              jitter_eps: 0.0
+              layer_1:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              layer_2:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              lr_scale: null
+              post_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              pre_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              recompute_level: full
+              router:
+                apply_peft: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              router_input_scale: 0.018844459036110227
+              router_normalization:
+                epsilon: 1.0e-06
+                lr_scale: null
+                type: fixed_rms_norm
+              router_per_expert_scale:
+                enabled: true
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              router_scale:
+                enabled: true
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              routing: aux_loss
+              shared_experts: 0
+              type: moe
+              z_loss_coefficient: 0.0
+            type: hybrid_moe
+          normalization:
+            epsilon: 1.0e-06
+            implementation: auto
+            lr_scale: null
+            type: rms_norm
+            weight:
+              initialization:
+                type: null
+              lr_scale: null
+              type: null
+            zero_centered: false
+          output_scale:
+            enabled: true
+            initialization:
+              type: null
+            lr_scale: 0.0
+            type: null
+          post_mixer_normalization:
+            epsilon: 1.0e-06
+            implementation: auto
+            lr_scale: null
+            type: rms_norm
+            weight:
+              initialization:
+                type: null
+              lr_scale: null
+              type: null
+            zero_centered: false
+          post_mlp_normalization:
+            epsilon: 1.0e-06
+            implementation: auto
+            lr_scale: null
+            type: rms_norm
+            weight:
+              initialization:
+                type: null
+              lr_scale: null
+              type: null
+            zero_centered: false
+          pre_mixer_normalization: null
+          pre_mlp_normalization:
+            lr_scale: null
+            type: none
+          type: decoder
+        sliding_attention:
+          distillation_loss_weight: 1.0
+          distillation_model: null
+          dropout: 0.0
+          lr_scale: null
+          mixer:
+            add_linear_biases: false
+            causal: true
+            dense_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            dropout: 0.0
+            head_groups: 8
+            head_size: 256
+            heads: 16
+            implementation: flash
+            key_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            key_norm:
+              epsilon: 1.0e-06
+              implementation: auto
+              lr_scale: null
+              type: rms_norm
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              zero_centered: false
+            lr_scale: null
+            query_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            query_norm:
+              epsilon: 1.0e-06
+              implementation: auto
+              lr_scale: null
+              type: rms_norm
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              zero_centered: false
+            rotary:
+              theta: 10000.0
+              type: default
+            shared_key_value: false
+            softmax_scale_power: 0.0
+            type: attention
+            use_flash_attention: true
+            value_layer:
+              apply_peft: null
+              bias:
+                enabled: null
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              lr_scale: null
+              type: null
+              weight:
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+            value_norm:
+              epsilon: 1.0e-06
+              lr_scale: null
+              type: fixed_rms_norm
+            window_size: 1024
+          mlp:
+            dense:
+              activation: gelu
+              add_linear_biases: false
+              gated: true
+              intermediate_size: 2112
+              layer_1:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              layer_2:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              lr_scale: null
+              post_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              pre_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              recompute_level: full
+              type: mlp
+            lr_scale: null
+            post_norm: null
+            pre_norm: null
+            routed:
+              activation: gelu
+              add_linear_biases: false
+              auxiliary_loss_coefficient: 0.0
+              dropless_dynamic_shape: false
+              experts: 128
+              experts_per_token: 8
+              gated: true
+              implementation: auto
+              intermediate_size: 704
+              jitter_eps: 0.0
+              layer_1:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              layer_2:
+                apply_peft: null
+                bias:
+                  enabled: null
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              lr_scale: null
+              post_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              pre_norm:
+                epsilon: 1.0e-06
+                implementation: auto
+                lr_scale: null
+                type: rms_norm
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+                zero_centered: false
+              recompute_level: full
+              router:
+                apply_peft: null
+                lr_scale: null
+                type: null
+                weight:
+                  initialization:
+                    type: null
+                  lr_scale: null
+                  type: null
+              router_input_scale: 0.018844459036110227
+              router_normalization:
+                epsilon: 1.0e-06
+                lr_scale: null
+                type: fixed_rms_norm
+              router_per_expert_scale:
+                enabled: true
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              router_scale:
+                enabled: true
+                initialization:
+                  type: null
+                lr_scale: null
+                type: null
+              routing: aux_loss
+              shared_experts: 0
+              type: moe
+              z_loss_coefficient: 0.0
+            type: hybrid_moe
+          normalization:
+            epsilon: 1.0e-06
+            implementation: auto
+            lr_scale: null
+            type: rms_norm
+            weight:
+              initialization:
+                type: null
+              lr_scale: null
+              type: null
+            zero_centered: false
+          output_scale:
+            enabled: true
+            initialization:
+              type: null
+            lr_scale: 0.0
+            type: null
+          post_mixer_normalization:
+            epsilon: 1.0e-06
+            implementation: auto
+            lr_scale: null
+            type: rms_norm
+            weight:
+              initialization:
+                type: null
+              lr_scale: null
+              type: null
+            zero_centered: false
+          post_mlp_normalization:
+            epsilon: 1.0e-06
+            implementation: auto
+            lr_scale: null
+            type: rms_norm
+            weight:
+              initialization:
+                type: null
+              lr_scale: null
+              type: null
+            zero_centered: false
+          pre_mixer_normalization: null
+          pre_mlp_normalization:
+            lr_scale: null
+            type: none
+          type: decoder
+      lr_scale: null
+      num_blocks: 30
+      pattern:
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - sliding_attention
+      - full_attention
+      type: pattern
+    embeddings:
+      dropout: 0.0
+      embedding_scale: 53.0659966456864
+      full_precision_residual: false
+      lr_scale: null
+      num_position_embeddings: 2048
+      position_embeddings:
+        enabled: null
+        initialization:
+          type: null
+        lr_scale: null
+        type: null
+      type: null
+      vocab_parallel: false
+      vocab_size: 262144
+      word_embeddings:
+        initialization:
+          type: null
+        lr_scale: null
+        type: null
+    head:
+      cross_entropy_splits: 8
+      final_logit_softcap: 30.0
+      logits_scale_factor: 1.0
+      losses: {}
+      lr_scale: null
+      normalization:
+        epsilon: 1.0e-06
+        implementation: auto
+        lr_scale: null
+        type: rms_norm
+        weight:
+          initialization:
+            type: null
+          lr_scale: null
+          type: null
+        zero_centered: false
+      output_weight:
+        initialization:
+          type: null
+        lr_scale: null
+        type: null
+      prediction_heads: 1
+      prediction_loss_coefficient: null
+      type: null
+    hidden_size: 2816
+    lr_scale: null
+    peft:
+      type: none
+    tied_embedding_weight: true
+    type: null
+    use_megatron_initialization: false
+  distributed:
+    backend: nccl
+    compute_dtype: bfloat16
+    dp_seed_shift: 317767863445754100075399033823
+    force_cpu_initialization: false
+    inference_seed_shift: 220111337975202516901860145957
+    local_world_size: 8
+    optimization_dtype: float32
+    pipeline_first: false
+    pipeline_parallel: 1
+    pp_gen_init_seed_shift: 631112027069964424381239824623
+    pp_gen_seed_shift: 500460795349110443334056239993
+    pp_seed_shift: 811026271858507938190098775099
+    rank: 0
+    reproducible_init: false
+    sample_seed_shift: 751127116949963770353413160199
+    seed: 1234
+    sequence_data_parallel: 4
+    sequence_tensor_parallel: true
+    tensor_parallel: 2
+    timeout: 3600.0
+    tp_gen_init_seed_shift: 894750739684993243926471979237
+    tp_gen_seed_shift: 278779420836085904093221202933
+    tp_seed_shift: 705275193289568515128435800471
+    train_seed_shift: 938219878163699459065752841447
+    type: null
+    use_cuda: true
+    valid_seed_shift: 683552447587140661489672773353
+    world_size: 32
+  multi_stage:
+    debug_activation_memory: false
+    debug_all_param_gradients: 0
+    debug_global_tensors: true
+    debug_layer_gradients: 0
+    debug_layer_outputs: 0
+    debug_param_gradients: 0
+    debug_param_init: 0
+    debug_param_update: 0
+    debug_tensor_parallel: false
+    full_precision_gradients: true
+    layers_per_stage: 1.0
+    num_grad_buffers: 2
+    num_weight_buffers: 2
+    pipeline_delay: 0.0
+    store_frozen_weights_in_optimization_precision: true
+    type: null
+    zero_stage: 3
+  type: gpt
+optimizer:
+  beta_1: 0.9
+  beta_2: 0.999
+  default_learning_rate_scale: 1.0
+  epsilon: 1.0e-08
+  gradient_norm_clipping: 1.0
+  gradient_scaler:
+    constant: null
+    hysteresis: 2
+    initial: 65536.0
+    minimum: 1.0
+    type: null
+    window: 1000
+  learning_rate:
+    base: 2.0e-05
+    decay_iterations: 210
+    decay_power: 1.0
+    decay_style: cosine
+    minimum: 0.0
+    type: null
+    warmup_iterations: 7
+  type: null
+  weight_decay: 0.01
+pretrained:
+  format: gemma4
+  load_config: model
+  model_weights: true
+  optimizer_state: false
+  path: /mnt/queue1/shruthan/models/gemma-4-26B-A4B-it-causal-lm
+  timeout: null
+  type: null
+profiling:
+  averages: false
+  cpu: false
+  cuda: false
+  cycles: 1
+  export: false
+  log: false
+  ranks: []
+  skip: 1
+  table_width: 80
+  trace: false
+  type: null
+  wait: 0
+  warmup: 1
+reference_models: {}
+run:
+  enable_all_loggers: false
+  enable_triton_kernels: true
+  experiment_dir: /mnt/queue1/shashank/exp/sft/26_09_08_sft_email_v4_v3
+  experiment_name: null
+  log_timestamps: true
+  model_debug_level: 0
+  structured_logs: true
+  tensor_logs:
+    full_tensors: false
+    max_elements: 8
+    save: false
+    show: true
+    type: null
+  torch_dynamo_enable: true
+  triton_linear_kernels: false
+  type: null
+schedule:
+  breadth_first_micro_batches: 1
+  data_batch_warn_time_ms: 1000.0
+  data_overlap: true
+  debug_schedule: false
+  debug_send_recv: false
+  depth_first_micro_batches: 16
+  micro_batch_splits: 1
+  pipeline_overlap: true
+  profile_schedule: false
+  skip_step: false
+  throttle_cpu: true
+  throttle_cpu_delay: 1
+  throttle_cpu_rate: 1
+  type: null
+training:
+  checkpoint:
+    interval: null
+    keep: 5
+    keep_every: null
+    offset: 0
+    type: null
+  evaluators: {}
+  export:
+    data_type: null
+    format: gemma4
+    interval: 70
+    keep: null
+    keep_every: null
+    model_weights: true
+    offset: 0
+    optimizer_state: false
+    parameters_per_file: 4294967296
+    type: null
+  logs:
+    interval: 1
+    offset: 0
+    type: null
+  num_workers: 0
+  prefetch_factor: null
+  shutdown:
+    interval: null
+    offset: 0
+    type: null
+  timeout: 3600.0
+  train_iters: 210
+  type: null
+  wandb:
+    alert:
+      interval: null
+      offset: 0
+      status_updates: null
+      type: null
+    entity_name: corellm
+    group_name: default
+    project_name: gemma-sft
+    type: null
+type: train_gpt

26_09_08_sft_email_v4_v3/configure_fastllm_model_config.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Fill in the data-dependent training knobs of a fastllm model_config.yaml.
+
+DATASETS (below) lists the sources to blend, each a `fast_llm_config.yaml` written by the
+`fast-llm prepare gpt_memmap` tokenization step (that file carries the source's real
+`metadata.num_tokens`). This script writes that list into model_config.yaml's
+data.datasets.training.datasets and derives the values that depend on their token counts:
+
+  data.datasets.training.datasets          DATASETS, as {type: file, path: ...} entries
+  data.datasets.training.weights           token count of each source (token-proportional blend)
+  schedule.depth_first_micro_batches        solved from TARGET_GLOBAL_BATCH_SIZE / NUM_NODES
+  training.train_iters                      NUM_EPOCHS * num_steps_per_epoch
+  training.checkpoint.interval              num_steps_per_epoch
+  training.export.interval                  num_steps_per_epoch
+  optimizer.learning_rate.warmup_iterations num_steps_per_epoch // 10
+  optimizer.learning_rate.decay_iterations  training.train_iters
+
+where, with TP=2 and SDP=4 already fixed in the model config, a single node contributes
+`depth_first_micro_batches` sequences per optimizer step (SDP/TP split one sequence across
+ranks rather than processing several in parallel), so:
+
+  global_batch_size    = NUM_NODES * depth_first_micro_batches
+  num_steps_per_epoch  = ceil(EXTRA_STEPS_FRACTION-padded total_tokens / (global_batch_size * seq_len)),
+                         then rounded up to the next multiple of 10
+
+seq_len is read from the model config's data.maximum_document_length. total_tokens is the sum
+of every source's num_tokens. The 5% padding accounts for packing-efficiency slack -- exact
+step counts are not expected to land precisely on an epoch boundary.
+
+All run configuration lives in the CONFIGURATION block below -- edit it directly rather than
+passing command-line arguments.
+"""
+
+from __future__ import annotations
+
+import math
+from pathlib import Path
+
+import yaml
+
+# -----------------------------------------------------------------------------
+# Configuration -- edit these instead of passing CLI arguments.
+# -----------------------------------------------------------------------------
+
+MODEL_CONFIG_PATH = Path("/mnt/queue1/shashank/exp/sft/26_09_08_sft_email_v4_v3/model_config.yaml")
+
+# fast_llm_config.yaml path for each source to blend into training
+DATASETS = [
+    "/mnt/queue1/shashank/exp/data/26_09_04_sft_email_v3/fastllm/fast_llm_config.yaml",
+    "/mnt/queue1/shashank/exp/data/26_09_08_sft_email_v4/fastllm/fast_llm_config.yaml",
+    "/mnt/queue1/shruthan/data/gemma_sft_replay/10k/fastllm/fast_llm_config.yaml",
+]
+
+NUM_NODES = 4
+TARGET_GLOBAL_BATCH_SIZE = 64
+NUM_EPOCHS = 3
+EXTRA_STEPS_FRACTION = 0.05  # padding for packing-efficiency slack
+
+
+def load_yaml(path: Path) -> dict:
+    with path.open("r", encoding="utf-8") as f:
+        return yaml.safe_load(f)
+
+
+def source_num_tokens(fast_llm_config_path: str) -> int:
+    source_config = load_yaml(Path(fast_llm_config_path))
+    return source_config["metadata"]["num_tokens"]
+
+
+def main() -> int:
+    config = load_yaml(MODEL_CONFIG_PATH)
+
+    seq_len = config["data"]["maximum_document_length"]
+
+    token_counts = [source_num_tokens(path) for path in DATASETS]
+    total_tokens = sum(token_counts)
+
+    depth_first_micro_batches = max(1, round(TARGET_GLOBAL_BATCH_SIZE / NUM_NODES))
+    global_batch_size = NUM_NODES * depth_first_micro_batches
+
+    num_steps_per_epoch = math.ceil((1.0 + EXTRA_STEPS_FRACTION) * total_tokens / (global_batch_size * seq_len))
+    num_steps_per_epoch = math.ceil(num_steps_per_epoch / 10) * 10
+    train_iters = NUM_EPOCHS * num_steps_per_epoch
+    warmup_iterations = num_steps_per_epoch // 10
+
+    config["data"]["datasets"]["training"]["datasets"] = [{"type": "file", "path": path} for path in DATASETS]
+    config["data"]["datasets"]["training"]["weights"] = token_counts
+    config["schedule"]["depth_first_micro_batches"] = depth_first_micro_batches
+    config["training"]["train_iters"] = train_iters
+    if "checkpoint" in config["training"]:
+        config["training"]["checkpoint"]["interval"] = num_steps_per_epoch
+    config["training"]["export"]["interval"] = num_steps_per_epoch
+    config["optimizer"]["learning_rate"]["warmup_iterations"] = warmup_iterations
+    config["optimizer"]["learning_rate"]["decay_iterations"] = train_iters
+
+    with MODEL_CONFIG_PATH.open("w", encoding="utf-8") as f:
+        yaml.safe_dump(config, f, sort_keys=False)
+
+    print(f"weights: {token_counts}  (total_tokens={total_tokens}, seq_len={seq_len})")
+    print(f"global_batch_size={global_batch_size}  depth_first_micro_batches={depth_first_micro_batches}")
+    print(f"num_steps_per_epoch={num_steps_per_epoch}  train_iters={train_iters}  warmup_iterations={warmup_iterations}")
+    print(f"Updated {MODEL_CONFIG_PATH}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shashank__exp__data__26_09_04_sft_email_v3__fastllm__shard_0_0.fast_llm_dataset_ns_2019_sl_32768_s_784569_shuffled_cumsum.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f78708060f6aca29e7c433ff174ad0b6a4c9d233f2ced9e56368a030a0ad1ea
+size 10504

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shashank__exp__data__26_09_04_sft_email_v3__fastllm__shard_0_0.fast_llm_dataset_ns_2019_sl_32768_s_784569_shuffling.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dddd9262b742465744ed1deceacdbf3468f0d932dd78c7e2eba3e5ccb4d7c14
+size 26088

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shashank__exp__data__26_09_04_sft_email_v3__fastllm__shard_0_0.yaml

@@ -0,0 +1,20 @@
+config:
+  dataset_name: training
+  gpu: true
+  maximum_document_length: 32768
+  micro_batch_size: 32768
+  predicted_tokens: 1
+  shuffle: epoch
+  token_cumsum_rate: 10
+  truncate_documents: false
+  type: null
+  world_size: 32
+dataset:
+  documents_per_epoch: 3947
+  name: __mnt__queue1__shashank__exp__data__26_09_04_sft_email_v3__fastllm__shard_0_0.fast_llm_dataset
+  tokens_per_epoch: 17841257
+num_samples: 2019
+sequence_length: 32768
+truncate_documents: false
+unshuffled_epochs: 0
+unshuffled_tokens: 0

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shashank__exp__data__26_09_08_sft_email_v4__fastllm__shard_0_0.fast_llm_dataset_ns_4417_sl_32768_s_785266_shuffled_cumsum.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b9ad75b4d0f5efc77c639274e4a7b412579961751e10483af37177e20337117
+size 17920

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shashank__exp__data__26_09_08_sft_email_v4__fastllm__shard_0_0.fast_llm_dataset_ns_4417_sl_32768_s_785266_shuffling.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e96c8ba14574954551a37b3f1dd22b01ea54109124dc2369201e70ec7393a642
+size 44628

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shashank__exp__data__26_09_08_sft_email_v4__fastllm__shard_0_0.yaml

@@ -0,0 +1,20 @@
+config:
+  dataset_name: training
+  gpu: true
+  maximum_document_length: 32768
+  micro_batch_size: 32768
+  predicted_tokens: 1
+  shuffle: epoch
+  token_cumsum_rate: 10
+  truncate_documents: false
+  type: null
+  world_size: 32
+dataset:
+  documents_per_epoch: 7626
+  name: __mnt__queue1__shashank__exp__data__26_09_08_sft_email_v4__fastllm__shard_0_0.fast_llm_dataset
+  tokens_per_epoch: 43465496
+num_samples: 4417
+sequence_length: 32768
+truncate_documents: false
+unshuffled_epochs: 0
+unshuffled_tokens: 0

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shruthan__data__gemma_sft_replay__10k__fastllm__shard_0_0.fast_llm_dataset_ns_7008_sl_32768_s_785963_shuffled_cumsum.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a682d417f335f5a4d4337a84cd615ec9622ed55d00c80b7930e2454209420cf
+size 22688

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shruthan__data__gemma_sft_replay__10k__fastllm__shard_0_0.fast_llm_dataset_ns_7008_sl_32768_s_785963_shuffling.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a52f9c807bdc37b3c62c6b8ffcb6a56db72b332f03027dfd48fed767a441c39
+size 56548

26_09_08_sft_email_v4_v3/dataset_cache/__mnt__queue1__shruthan__data__gemma_sft_replay__10k__fastllm__shard_0_0.yaml

@@ -0,0 +1,20 @@
+config:
+  dataset_name: training
+  gpu: true
+  maximum_document_length: 32768
+  micro_batch_size: 32768
+  predicted_tokens: 1
+  shuffle: epoch
+  token_cumsum_rate: 10
+  truncate_documents: false
+  type: null
+  world_size: 32
+dataset:
+  documents_per_epoch: 10000
+  name: __mnt__queue1__shruthan__data__gemma_sft_replay__10k__fastllm__shard_0_0.fast_llm_dataset
+  tokens_per_epoch: 67917533
+num_samples: 7008
+sequence_length: 32768
+truncate_documents: false
+unshuffled_epochs: 0
+unshuffled_tokens: 0

26_09_08_sft_email_v4_v3/export/gemma4/140/chat_template.jinja

@@ -0,0 +1,390 @@
+{#
+  Template: Google Gemma 4 Canonical Chat Template
+  Author: Google Gemma Engineering Team
+  Published: 2026-07-09
+  Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
+#}
+{%- macro format_parameters(properties, required, filter_keys=false) -%}
+    {%- set standard_keys = ['properties', 'required', 'nullable'] -%}
+    {%- set ns = namespace(found_first=false) -%}
+    {%- for key, value in properties | dictsort -%}
+        {%- set add_comma = false -%}
+        {%- if not filter_keys or key not in standard_keys -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {{ key }}:{
+            {%- if value['description'] -%}
+                description:<|"|>{{ value['description'] }}<|"|>
+                {%- set add_comma = true -%}
+            {%- endif -%}
+            {%- if value['type'] | upper == 'STRING' -%}
+                {%- if value['enum'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    enum:{{ format_argument(value['enum']) }}
+                {%- endif -%}
+            {%- elif value['type'] | upper == 'ARRAY' -%}
+                {%- if value['items'] is mapping and value['items'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    items:{
+                    {%- set ns_items = namespace(found_first=false) -%}
+                    {%- for item_key, item_value in value['items'] | dictsort -%}
+                        {%- if item_value is not none -%}
+                            {%- if ns_items.found_first %},{% endif -%}
+                            {%- set ns_items.found_first = true -%}
+                            {%- if item_key == 'properties' -%}
+                                properties:{
+                                {%- if item_value is mapping -%}
+                                    {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
+                                {%- endif -%}
+                                }
+                            {%- elif item_key == 'required' -%}
+                                required:[
+                                {%- for req_item in item_value -%}
+                                    <|"|>{{- req_item -}}<|"|>
+                                    {%- if not loop.last %},{% endif -%}
+                                {%- endfor -%}
+                                ]
+                            {%- elif item_key == 'type' -%}
+                                {%- if item_value is string -%}
+                                    type:{{ format_argument(item_value | upper) }}
+                                {%- else -%}
+                                    type:{{ format_argument(item_value | map('upper') | list) }}
+                                {%- endif -%}
+                            {%- else -%}
+                                {{ item_key }}:{{ format_argument(item_value) }}
+                            {%- endif -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                    }
+                {%- endif -%}
+            {%- endif -%}
+            {%- if value['nullable'] %}
+                {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                nullable:true
+            {%- endif -%}
+            {%- if value['type'] | upper == 'OBJECT' -%}
+                {%- if value['properties'] is defined and value['properties'] is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value['properties'], value['required'] | default([])) -}}
+                    }
+                {%- elif value is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}
+                    }
+                {%- endif -%}
+                {%- if value['required'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    required:[
+                    {%- for item in value['required'] | default([]) -%}
+                        <|"|>{{- item -}}<|"|>
+                        {%- if not loop.last %},{% endif -%}
+                    {%- endfor -%}
+                    ]
+                {%- endif -%}
+            {%- endif -%}
+            {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+            type:<|"|>{{ value['type'] | upper }}<|"|>}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endmacro -%}
+{%- macro format_function_declaration(tool_data) -%}
+    declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
+    {%- set params = tool_data['function']['parameters'] -%}
+    {%- if params -%}
+        ,parameters:{
+        {%- if params['properties'] -%}
+            properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
+        {%- endif -%}
+        {%- if params['required'] -%}
+            required:[
+            {%- for item in params['required'] -%}
+                <|"|>{{- item -}}<|"|>
+                {{- ',' if not loop.last -}}
+            {%- endfor -%}
+            ],
+        {%- endif -%}
+        {%- if params['type'] -%}
+            type:<|"|>{{- params['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    {%- if 'response' in tool_data['function'] -%}
+        {%- set response_declaration = tool_data['function']['response'] -%}
+        ,response:{
+        {%- if response_declaration['description'] -%}
+            description:<|"|>{{- response_declaration['description'] -}}<|"|>,
+        {%- endif -%}
+        {%- if response_declaration['type'] | upper == 'OBJECT' -%}
+            type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    }
+{%- endmacro -%}
+{%- macro format_argument(argument, escape_keys=True) -%}
+    {%- if argument is none -%}
+        {{- 'null' -}}
+    {%- elif argument is string -%}
+        {{- '<|"|>' + argument + '<|"|>' -}}
+    {%- elif argument is boolean -%}
+        {{- 'true' if argument else 'false' -}}
+    {%- elif argument is mapping -%}
+        {{- '{' -}}
+        {%- set ns = namespace(found_first=false) -%}
+        {%- for key, value in argument | dictsort -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {%- if escape_keys -%}
+                {{- '<|"|>' + key + '<|"|>' -}}
+            {%- else -%}
+                {{- key -}}
+            {%- endif -%}
+            :{{- format_argument(value, escape_keys=escape_keys) -}}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- elif argument is sequence -%}
+        {{- '[' -}}
+        {%- for item in argument -%}
+            {{- format_argument(item, escape_keys=escape_keys) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- ']' -}}
+    {%- else -%}
+        {{- argument -}}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro strip_thinking(text) -%}
+    {%- set ns = namespace(result='') -%}
+    {%- for part in text.split('<channel|>') -%}
+        {%- if '<|channel>' in part -%}
+            {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
+        {%- else -%}
+            {%- set ns.result = ns.result + part -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {{- ns.result | trim -}}
+{%- endmacro -%}
+
+{%- macro format_tool_response_block(tool_name, response) -%}
+    {{- '<|tool_response>' -}}
+    {%- if response is mapping -%}
+        {{- 'response:' + tool_name + '{' -}}
+        {%- for key, value in response | dictsort -%}
+            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- else -%}
+        {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
+    {%- endif -%}
+    {{- '<tool_response|>' -}}
+{%- endmacro -%}
+
+{#- ===== SETUP ===== -#}
+{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
+{%- set loop_messages = messages -%}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
+{{- bos_token -}}
+{#- Handle System/Tool Definitions Block -#}
+{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
+    {{- '<|turn>system\n' -}}
+    {#- Inject Thinking token at the very top of the FIRST system turn -#}
+    {%- if enable_thinking -%}
+        {{- '<|think|>\n' -}}
+        {%- set ns.prev_message_type = 'think' -%}
+    {%- endif -%}
+    {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
+        {%- if messages[0]['content'] is string -%}
+            {{- messages[0]['content'] | trim -}}
+        {%- elif messages[0]['content'] is sequence -%}
+            {%- for item in messages[0]['content'] -%}
+                {{- item['text'] | trim + ' '-}}
+            {%- endfor -%}
+        {%- endif -%}
+        {%- set loop_messages = messages[1:] -%}
+    {%- endif -%}
+    {%- if tools -%}
+        {%- for tool in tools %}
+            {{- '<|tool>' -}}
+            {{- format_function_declaration(tool) | trim -}}
+            {{- '<tool|>' -}}
+        {%- endfor %}
+        {%- set ns.prev_message_type = 'tool' -%}
+    {%- endif -%}
+    {{- '<turn|>\n' -}}
+{%- endif %}
+
+{#- Pre-scan: find last user message index for reasoning guard -#}
+{%- set ns_turn = namespace(last_user_idx=-1) -%}
+{%- for i in range(loop_messages | length) -%}
+    {%- if loop_messages[i]['role'] == 'user' -%}
+        {%- set ns_turn.last_user_idx = i -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Loop through messages -#}
+{%- for message in loop_messages -%}
+    {%- if message['role'] != 'tool' -%}
+    {%- set ns.prev_message_type = None -%}
+    {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
+    {#- Detect continuation using tracked state — O(1) instead of O(n) backward scan -#}
+    {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
+    {%- if not continue_same_model_turn -%}
+        {{- '<|turn>' + role + '\n' }}
+
+    {%- endif -%}
+
+    {#- Render reasoning/reasoning_content as thinking channel -#}
+    {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
+    {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}
+    {%- if thinking_text and thinking_gate -%}
+        {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
+    {%- endif -%}
+
+            {%- if message.get('tool_calls') -%}
+                {%- for tool_call in message.get('tool_calls') -%}
+                    {%- set function = tool_call['function'] -%}
+                    {{- '<|tool_call>call:' + function['name'] + '{' -}}
+                    {%- if function['arguments'] is mapping -%}
+                        {%- set ns_args = namespace(found_first=false) -%}
+                        {%- for key, value in function['arguments'] | dictsort -%}
+                            {%- if ns_args.found_first %},{% endif -%}
+                            {%- set ns_args.found_first = true -%}
+                            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+                        {%- endfor -%}
+                    {%- elif function['arguments'] is none -%}
+                    {%- else -%}
+                        {{- raise_exception(
+                            "chat_template: tool_calls[].function.arguments must be a "
+                            "JSON object (mapping), not a string. Deserialize arguments "
+                            "before passing to the template."
+                        ) -}}
+                    {%- endif -%}
+                    {{- '}<tool_call|>' -}}
+                {%- endfor -%}
+                {%- set ns.prev_message_type = 'tool_call' -%}
+            {%- endif -%}
+
+            {%- set ns_tr_out = namespace(flag=false) -%}
+            {%- if message.get('tool_responses') -%}
+                {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
+                {%- for tool_response in message.get('tool_responses') -%}
+                    {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
+                    {%- set ns_tr_out.flag = true -%}
+                    {%- set ns.prev_message_type = 'tool_response' -%}
+                {%- endfor -%}
+            {%- elif message.get('tool_calls') -%}
+                {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
+                {%- set ns_tool_scan = namespace(stopped=false) -%}
+                {%- for k in range(loop.index0 + 1, loop_messages | length) -%}
+                    {%- if ns_tool_scan.stopped -%}
+                    {%- elif loop_messages[k]['role'] != 'tool' -%}
+                        {%- set ns_tool_scan.stopped = true -%}
+                    {%- else -%}
+                        {%- set follow = loop_messages[k] -%}
+                        {#- Resolve tool_call_id to function name -#}
+                        {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
+                        {%- for tc in message.get('tool_calls') -%}
+                            {%- if tc.get('id') == follow.get('tool_call_id') -%}
+                                {%- set ns_tname.name = tc['function']['name'] -%}
+                            {%- endif -%}
+                        {%- endfor -%}
+                        {#- Handle content as string or content-parts array -#}
+                        {%- set tool_body = follow.get('content') -%}
+                        {%- if tool_body is string -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- elif tool_body is sequence and tool_body is not string -%}
+                            {%- set ns_txt = namespace(s='') -%}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') == 'text' -%}
+                                    {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
+                                {%- endif -%}
+                            {%- endfor -%}
+                            {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') in ['image', 'image_url'] -%}
+                                    {{- '<|image|>' -}}
+                                {%- elif part.get('type') in ['audio', 'input_audio'] -%}
+                                    {{- '<|audio|>' -}}
+                                {%- elif part.get('type') == 'video' -%}
+                                    {{- '<|video|>' -}}
+                                {%- endif -%}
+                            {%- endfor -%}
+                        {%- else -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- endif -%}
+                        {%- set ns_tr_out.flag = true -%}
+                        {%- set ns.prev_message_type = 'tool_response' -%}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+
+            {%- set captured_content -%}
+            {%- if message.get('content') is string -%}
+                {%- if role == 'model' -%}
+                    {{- strip_thinking(message['content']) -}}
+                {%- else -%}
+                    {{- message['content'] | trim -}}
+                {%- endif -%}
+            {%- elif message.get('content') is sequence -%}
+                {%- for item in message['content'] -%}
+                    {%- if item.get('type') == 'text' -%}
+                        {%- if role == 'model' -%}
+                            {{- strip_thinking(item['text']) -}}
+                        {%- else -%}
+                            {{- item['text'] | trim -}}
+                        {%- endif -%}
+                    {%- elif item.get('type') in ['image', 'image_url'] -%}
+                        {{- '<|image|>' -}}
+                    {%- elif item.get('type') in ['audio', 'input_audio'] -%}
+                        {{- '<|audio|>' -}}
+                    {%- elif item.get('type') == 'video' -%}
+                        {{- '<|video|>' -}}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+            {%- endset -%}
+
+            {{- captured_content -}}
+            {%- set has_content = captured_content | trim | length > 0 -%}
+
+        {#- Forward-scan: find next non-tool message role for continuation detection -#}
+        {%- set next_nt = namespace(role=None, found=false) -%}
+        {%- for j in range(loop.index0 + 1, loop_messages | length) -%}
+            {%- if not next_nt.found -%}
+                {%- if loop_messages[j]['role'] != 'tool' -%}
+                    {%- set next_nt.role = loop_messages[j]['role'] -%}
+                    {%- set next_nt.found = true -%}
+                {%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+
+        {%- set continues_into_next = (
+            role == 'model'
+            and next_nt.role == 'assistant'
+            and (not message.get('tool_calls') or ns_tr_out.flag)
+        ) -%}
+
+        {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
+            {{- '<|tool_response>' -}}
+        {%- elif continues_into_next -%}
+        {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}
+            {{- '<turn|>\n' -}}
+        {%- endif -%}
+
+    {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
+    {%- set ns.prev_non_tool_role = message['role'] -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if add_generation_prompt -%}
+    {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
+        {{- '<|turn>model\n' -}}
+        {%- if not enable_thinking -%}
+            {{- '<|channel>thought\n<channel|>' -}}
+        {%- endif -%}
+    {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
+        {{- '<|channel>thought\n' -}}
+    {%- endif -%}
+{%- endif -%}

26_09_08_sft_email_v4_v3/export/gemma4/140/config.json

@@ -0,0 +1,83 @@
+{
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attention_k_eq_v": true,
+  "bos_token_id": 2,
+  "dtype": "bfloat16",
+  "enable_moe_block": true,
+  "eos_token_id": 1,
+  "final_logit_softcapping": 30.0,
+  "global_head_dim": 512,
+  "head_dim": 256,
+  "hidden_activation": "gelu_pytorch_tanh",
+  "hidden_size": 2816,
+  "hidden_size_per_layer_input": 0,
+  "initializer_range": 0.02,
+  "intermediate_size": 2112,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 262144,
+  "model_type": "gemma4_text",
+  "moe_intermediate_size": 704,
+  "num_attention_heads": 16,
+  "num_experts": 128,
+  "num_global_key_value_heads": 2,
+  "num_hidden_layers": 30,
+  "num_key_value_heads": 8,
+  "num_kv_shared_layers": 0,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "full_attention": {
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 1000000.0,
+      "rope_type": "proportional"
+    },
+    "sliding_attention": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    }
+  },
+  "sliding_window": 1024,
+  "tie_word_embeddings": true,
+  "top_k_experts": 8,
+  "use_bidirectional_attention": "vision",
+  "use_cache": true,
+  "use_double_wide_mlp": false,
+  "vocab_size": 262144,
+  "vocab_size_per_layer_input": 262144,
+  "architectures": [
+    "Gemma4ForCausalLM"
+  ],
+  "transformers_version": "5.5.0.dev0"
+}

26_09_08_sft_email_v4_v3/export/gemma4/140/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 2,
+  "do_sample": true,
+  "eos_token_id": [
+    1,
+    106,
+    50
+  ],
+  "pad_token_id": 0,
+  "temperature": 1.0,
+  "top_k": 64,
+  "top_p": 0.95,
+  "transformers_version": "5.5.0.dev0"
+}

26_09_08_sft_email_v4_v3/export/gemma4/140/model.safetensors.index.json

@@ -0,0 +1,1053 @@
+{
+    "metadata": {
+        "fast_llm_metadata": {
+            "fast_llm_version": "0.4.0",
+            "model": "gpt",
+            "format": "gemma4",
+            "config": {
+                "type": "gpt",
+                "base_model": {
+                    "decoder": {
+                        "type": "pattern",
+                        "blocks": {
+                            "full_attention": {
+                                "type": "decoder",
+                                "mixer": {
+                                    "type": "attention",
+                                    "rotary": {
+                                        "type": "proportional",
+                                        "theta": 1000000.0,
+                                        "partial_rotary_factor": 0.25
+                                    },
+                                    "heads": 16,
+                                    "head_groups": 2,
+                                    "head_size": 512,
+                                    "add_linear_biases": false,
+                                    "dropout": 0.0,
+                                    "softmax_scale_power": 0.0,
+                                    "implementation": "sdpa",
+                                    "query_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "key_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "value_norm": {
+                                        "type": "fixed_rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "shared_key_value": true
+                                },
+                                "mlp": {
+                                    "type": "hybrid_moe",
+                                    "dense": {
+                                        "type": "mlp",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 2112,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full"
+                                    },
+                                    "routed": {
+                                        "type": "moe",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 704,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full",
+                                        "router_normalization": {
+                                            "type": "fixed_rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "router_scale": {
+                                            "enabled": true
+                                        },
+                                        "router_input_scale": 0.018844459036110227,
+                                        "router_per_expert_scale": {
+                                            "enabled": true
+                                        },
+                                        "experts": 128,
+                                        "experts_per_token": 8,
+                                        "auxiliary_loss_coefficient": 0.0
+                                    }
+                                },
+                                "normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "pre_mlp_normalization": {
+                                    "type": "none"
+                                },
+                                "post_mixer_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "post_mlp_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "output_scale": {
+                                    "lr_scale": 0.0,
+                                    "enabled": true
+                                }
+                            },
+                            "sliding_attention": {
+                                "type": "decoder",
+                                "mixer": {
+                                    "type": "attention",
+                                    "rotary": {
+                                        "type": "default",
+                                        "theta": 10000.0
+                                    },
+                                    "heads": 16,
+                                    "head_groups": 8,
+                                    "head_size": 256,
+                                    "add_linear_biases": false,
+                                    "dropout": 0.0,
+                                    "window_size": 1024,
+                                    "softmax_scale_power": 0.0,
+                                    "implementation": "flash",
+                                    "query_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "key_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "value_norm": {
+                                        "type": "fixed_rms_norm",
+                                        "epsilon": 1e-06
+                                    }
+                                },
+                                "mlp": {
+                                    "type": "hybrid_moe",
+                                    "dense": {
+                                        "type": "mlp",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 2112,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full"
+                                    },
+                                    "routed": {
+                                        "type": "moe",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 704,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full",
+                                        "router_normalization": {
+                                            "type": "fixed_rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "router_scale": {
+                                            "enabled": true
+                                        },
+                                        "router_input_scale": 0.018844459036110227,
+                                        "router_per_expert_scale": {
+                                            "enabled": true
+                                        },
+                                        "experts": 128,
+                                        "experts_per_token": 8,
+                                        "auxiliary_loss_coefficient": 0.0
+                                    }
+                                },
+                                "normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "pre_mlp_normalization": {
+                                    "type": "none"
+                                },
+                                "post_mixer_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "post_mlp_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "output_scale": {
+                                    "lr_scale": 0.0,
+                                    "enabled": true
+                                }
+                            }
+                        },
+                        "pattern": [
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention"
+                        ],
+                        "num_blocks": 30
+                    },
+                    "embeddings": {
+                        "vocab_size": 262144,
+                        "vocab_parallel": false,
+                        "embedding_scale": 53.0659966456864
+                    },
+                    "head": {
+                        "normalization": {
+                            "type": "rms_norm",
+                            "epsilon": 1e-06
+                        },
+                        "cross_entropy_splits": 8,
+                        "final_logit_softcap": 30.0
+                    },
+                    "tied_embedding_weight": true,
+                    "hidden_size": 2816,
+                    "peft": {
+                        "type": "none"
+                    }
+                },
+                "multi_stage": {
+                    "zero_stage": 3
+                },
+                "distributed": {
+                    "tensor_parallel": 2,
+                    "sequence_tensor_parallel": true,
+                    "sequence_data_parallel": 4,
+                    "world_size": 32,
+                    "rank": 0,
+                    "local_world_size": 8,
+                    "timeout": 3600.0,
+                    "compute_dtype": "bfloat16"
+                }
+            },
+            "shards": [
+                "weights"
+            ],
+            "metadata": {
+                "optimizer": {
+                    "current_step": 140,
+                    "grad_scaler": {
+                        "type": "NoopGradScaler"
+                    }
+                },
+                "completed_steps": 140,
+                "metrics": {
+                    "training": {
+                        "batch_size": 2097152,
+                        "load_balancing_loss": 828.73388671875,
+                        "lm_head_loss": 0.11952575296163559,
+                        "total_steps": 210,
+                        "completed_steps": 140,
+                        "consumed_tokens": 293601280,
+                        "percent_done": 66.66666666666667,
+                        "step_time_ms": 26989.178344607353,
+                        "step_time_average_ms": 30017.633229094958,
+                        "remaining_time": 2101.234326036647,
+                        "completion_time": 1788873030.2734404,
+                        "skipped_iters": 0,
+                        "nan_iters": 0,
+                        "model_tflops": 62205.328686694214,
+                        "hardware_tflops": 77174.52198293927,
+                        "tokens_per_sec_per_gpu": 2428.232499826902,
+                        "run": 0,
+                        "grad_norm": 2.5687570571899414,
+                        "learning_rate": 5.3159155930020984e-06,
+                        "loss_scale": 1.0,
+                        "reserved": 55562.0,
+                        "allocated": 21478.49755859375,
+                        "max_allocated": 54799.8583984375,
+                        "max_reserved": 55562.0,
+                        "global_max_reserved": 74782.0
+                    }
+                }
+            }
+        },
+        "model_config": {
+            "vocab_size": 262144,
+            "num_attention_heads": 16,
+            "num_key_value_heads": 8,
+            "head_dim": 256,
+            "global_head_dim": 512,
+            "num_global_key_value_heads": 2,
+            "attention_bias": false,
+            "attention_dropout": 0.0,
+            "sliding_window": 1024,
+            "rms_norm_eps": 1e-06,
+            "attention_k_eq_v": true,
+            "rope_parameters": {
+                "sliding_attention": {
+                    "rope_type": "default",
+                    "rope_theta": 10000.0
+                },
+                "full_attention": {
+                    "rope_type": "proportional",
+                    "rope_theta": 1000000.0,
+                    "partial_rotary_factor": 0.25
+                }
+            },
+            "intermediate_size": 2112,
+            "hidden_activation": "gelu_pytorch_tanh",
+            "num_experts": 128,
+            "top_k_experts": 8,
+            "moe_intermediate_size": 704,
+            "enable_moe_block": true,
+            "num_hidden_layers": 30,
+            "layer_types": [
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention"
+            ],
+            "final_logit_softcapping": 30.0,
+            "tie_word_embeddings": true,
+            "hidden_size": 2816,
+            "hidden_size_per_layer_input": 0,
+            "use_bidirectional_attention": null,
+            "model_type": "gemma4_text",
+            "architectures": [
+                "Gemma4ForCausalLM"
+            ]
+        },
+        "format": "pt"
+    },
+    "weight_map": {
+        "model.embed_tokens.weight": "model_0.safetensors",
+        "model.layers.0.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.0.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.0.router.scale": "model_0.safetensors",
+        "model.layers.0.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.0.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.0.router.proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.0.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.0.experts.down_proj": "model_0.safetensors",
+        "model.layers.0.layer_scalar": "model_0.safetensors",
+        "model.layers.1.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.1.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.1.router.scale": "model_0.safetensors",
+        "model.layers.1.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.1.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.1.router.proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.1.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.1.experts.down_proj": "model_0.safetensors",
+        "model.layers.1.layer_scalar": "model_0.safetensors",
+        "model.layers.2.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.2.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.2.router.scale": "model_0.safetensors",
+        "model.layers.2.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.2.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.2.router.proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.2.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.2.experts.down_proj": "model_0.safetensors",
+        "model.layers.2.layer_scalar": "model_0.safetensors",
+        "model.layers.3.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.3.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.3.router.scale": "model_0.safetensors",
+        "model.layers.3.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.3.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.3.router.proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.3.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.3.experts.down_proj": "model_0.safetensors",
+        "model.layers.3.layer_scalar": "model_0.safetensors",
+        "model.layers.4.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.4.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.4.router.scale": "model_0.safetensors",
+        "model.layers.4.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.4.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.4.router.proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.4.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.4.experts.down_proj": "model_1.safetensors",
+        "model.layers.4.layer_scalar": "model_1.safetensors",
+        "model.layers.5.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.5.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.5.router.scale": "model_1.safetensors",
+        "model.layers.5.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.5.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.5.router.proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.5.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.5.experts.down_proj": "model_1.safetensors",
+        "model.layers.5.layer_scalar": "model_1.safetensors",
+        "model.layers.6.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.6.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.6.router.scale": "model_1.safetensors",
+        "model.layers.6.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.6.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.6.router.proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.6.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.6.experts.down_proj": "model_1.safetensors",
+        "model.layers.6.layer_scalar": "model_1.safetensors",
+        "model.layers.7.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.7.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.7.router.scale": "model_1.safetensors",
+        "model.layers.7.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.7.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.7.router.proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.7.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.7.experts.down_proj": "model_1.safetensors",
+        "model.layers.7.layer_scalar": "model_1.safetensors",
+        "model.layers.8.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.8.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.8.router.scale": "model_1.safetensors",
+        "model.layers.8.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.8.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.8.router.proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.8.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.8.experts.down_proj": "model_1.safetensors",
+        "model.layers.8.layer_scalar": "model_1.safetensors",
+        "model.layers.9.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.9.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.9.router.scale": "model_1.safetensors",
+        "model.layers.9.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.9.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.9.router.proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.9.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.9.experts.down_proj": "model_1.safetensors",
+        "model.layers.9.layer_scalar": "model_2.safetensors",
+        "model.layers.10.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.10.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.10.router.scale": "model_2.safetensors",
+        "model.layers.10.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.10.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.10.router.proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.10.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.10.experts.down_proj": "model_2.safetensors",
+        "model.layers.10.layer_scalar": "model_2.safetensors",
+        "model.layers.11.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.11.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.11.router.scale": "model_2.safetensors",
+        "model.layers.11.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.11.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.11.router.proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.11.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.11.experts.down_proj": "model_2.safetensors",
+        "model.layers.11.layer_scalar": "model_2.safetensors",
+        "model.layers.12.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.12.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.12.router.scale": "model_2.safetensors",
+        "model.layers.12.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.12.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.12.router.proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.12.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.12.experts.down_proj": "model_2.safetensors",
+        "model.layers.12.layer_scalar": "model_2.safetensors",
+        "model.layers.13.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.13.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.13.router.scale": "model_2.safetensors",
+        "model.layers.13.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.13.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.13.router.proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.13.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.13.experts.down_proj": "model_2.safetensors",
+        "model.layers.13.layer_scalar": "model_2.safetensors",
+        "model.layers.14.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.14.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.14.router.scale": "model_2.safetensors",
+        "model.layers.14.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.14.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.14.router.proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.14.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.14.experts.down_proj": "model_2.safetensors",
+        "model.layers.14.layer_scalar": "model_2.safetensors",
+        "model.layers.15.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.15.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.15.router.scale": "model_2.safetensors",
+        "model.layers.15.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.15.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.15.router.proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.15.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.15.experts.down_proj": "model_3.safetensors",
+        "model.layers.15.layer_scalar": "model_3.safetensors",
+        "model.layers.16.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.16.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.16.router.scale": "model_3.safetensors",
+        "model.layers.16.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.16.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.16.router.proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.16.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.16.experts.down_proj": "model_3.safetensors",
+        "model.layers.16.layer_scalar": "model_3.safetensors",
+        "model.layers.17.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.17.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.17.router.scale": "model_3.safetensors",
+        "model.layers.17.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.17.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.17.router.proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.17.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.17.experts.down_proj": "model_3.safetensors",
+        "model.layers.17.layer_scalar": "model_3.safetensors",
+        "model.layers.18.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.18.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.18.router.scale": "model_3.safetensors",
+        "model.layers.18.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.18.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.18.router.proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.18.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.18.experts.down_proj": "model_3.safetensors",
+        "model.layers.18.layer_scalar": "model_3.safetensors",
+        "model.layers.19.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.19.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.19.router.scale": "model_3.safetensors",
+        "model.layers.19.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.19.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.19.router.proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.19.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.19.experts.down_proj": "model_3.safetensors",
+        "model.layers.19.layer_scalar": "model_3.safetensors",
+        "model.layers.20.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.20.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.20.router.scale": "model_3.safetensors",
+        "model.layers.20.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.20.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.20.router.proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.20.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.20.experts.down_proj": "model_3.safetensors",
+        "model.layers.20.layer_scalar": "model_4.safetensors",
+        "model.layers.21.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.21.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.21.router.scale": "model_4.safetensors",
+        "model.layers.21.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.21.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.21.router.proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.21.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.21.experts.down_proj": "model_4.safetensors",
+        "model.layers.21.layer_scalar": "model_4.safetensors",
+        "model.layers.22.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.22.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.22.router.scale": "model_4.safetensors",
+        "model.layers.22.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.22.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.22.router.proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.22.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.22.experts.down_proj": "model_4.safetensors",
+        "model.layers.22.layer_scalar": "model_4.safetensors",
+        "model.layers.23.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.23.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.23.router.scale": "model_4.safetensors",
+        "model.layers.23.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.23.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.23.router.proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.23.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.23.experts.down_proj": "model_4.safetensors",
+        "model.layers.23.layer_scalar": "model_4.safetensors",
+        "model.layers.24.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.24.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.24.router.scale": "model_4.safetensors",
+        "model.layers.24.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.24.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.24.router.proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.24.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.24.experts.down_proj": "model_4.safetensors",
+        "model.layers.24.layer_scalar": "model_4.safetensors",
+        "model.layers.25.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.25.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.25.router.scale": "model_4.safetensors",
+        "model.layers.25.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.25.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.25.router.proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.25.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.25.experts.down_proj": "model_4.safetensors",
+        "model.layers.25.layer_scalar": "model_4.safetensors",
+        "model.layers.26.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.26.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.26.router.scale": "model_4.safetensors",
+        "model.layers.26.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.26.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.26.router.proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.26.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.26.experts.down_proj": "model_5.safetensors",
+        "model.layers.26.layer_scalar": "model_5.safetensors",
+        "model.layers.27.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.27.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.27.router.scale": "model_5.safetensors",
+        "model.layers.27.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.27.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.27.router.proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.v_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.27.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.27.experts.down_proj": "model_5.safetensors",
+        "model.layers.27.layer_scalar": "model_5.safetensors",
+        "model.layers.28.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.28.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.28.router.scale": "model_5.safetensors",
+        "model.layers.28.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.28.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.28.router.proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.v_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.28.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.28.experts.down_proj": "model_5.safetensors",
+        "model.layers.28.layer_scalar": "model_5.safetensors",
+        "model.layers.29.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.29.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.29.router.scale": "model_5.safetensors",
+        "model.layers.29.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.29.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.29.router.proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.29.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.29.experts.down_proj": "model_5.safetensors",
+        "model.layers.29.layer_scalar": "model_5.safetensors",
+        "model.norm.weight": "model_5.safetensors"
+    }
+}

26_09_08_sft_email_v4_v3/export/gemma4/140/model_0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7277a2f42a87e19765332e8d65f5f57a4fc8cd9f4fa0dbaa4c48f5b110544a13
+size 18219689440

26_09_08_sft_email_v4_v3/export/gemma4/140/model_1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:204e8a7fe92c63083155d33933c1a6da90188f699f6e0a9836bef21c3b14098c
+size 17354616396

26_09_08_sft_email_v4_v3/export/gemma4/140/model_2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6230bf7ac58d23004a10d04b2a8a02da00323c198337ade4ab69e97bb9355f02
+size 18580955576

26_09_08_sft_email_v4_v3/export/gemma4/140/model_3.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d3a2cad124dd0352a5285efc72021131105b10f3fe8d1c2eaa8ab988da7133e
+size 17354616508

26_09_08_sft_email_v4_v3/export/gemma4/140/model_4.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49322496e747f7c8f2535553d59da3efef60dc07cd9181b908c6d3515b1c00eb
+size 18580955560

26_09_08_sft_email_v4_v3/export/gemma4/140/model_5.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:838deccb7a04bf4e7b1dcbf8eaf043dc0b0dbbb2b775073e19e1f352ee86aaba
+size 10841863008

26_09_08_sft_email_v4_v3/export/gemma4/140/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f
+size 32169626

26_09_08_sft_email_v4_v3/export/gemma4/140/tokenizer_config.json

@@ -0,0 +1,74 @@
+{
+  "audio_token": "<|audio|>",
+  "backend": "tokenizers",
+  "boa_token": "<|audio>",
+  "boi_token": "<|image>",
+  "bos_token": "<bos>",
+  "eoa_token": "<audio|>",
+  "eoc_token": "<channel|>",
+  "eoi_token": "<image|>",
+  "eos_token": "<eos>",
+  "eot_token": "<turn|>",
+  "escape_token": "<|\"|>",
+  "etc_token": "<tool_call|>",
+  "etd_token": "<tool|>",
+  "etr_token": "<tool_response|>",
+  "extra_special_tokens": [
+    "<|video|>"
+  ],
+  "image_token": "<|image|>",
+  "mask_token": "<mask>",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "padding_side": "left",
+  "processor_class": "Gemma4Processor",
+  "response_schema": {
+    "type": "object",
+    "properties": {
+      "role": {
+        "const": "assistant"
+      },
+      "thinking": {
+        "type": "string"
+      },
+      "content": {
+        "type": "string"
+      },
+      "tool_calls": {
+        "x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>",
+        "type": "array",
+        "items": {
+          "type": "object",
+          "properties": {
+            "type": {
+              "const": "function"
+            },
+            "function": {
+              "type": "object",
+              "x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})",
+              "properties": {
+                "name": {
+                  "type": "string"
+                },
+                "arguments": {
+                  "type": "object",
+                  "x-parser": "gemma4-tool-call",
+                  "additionalProperties": {}
+                }
+              }
+            }
+          }
+        }
+      }
+    },
+    "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
+  },
+  "soc_token": "<|channel>",
+  "sot_token": "<|turn>",
+  "stc_token": "<|tool_call>",
+  "std_token": "<|tool>",
+  "str_token": "<|tool_response>",
+  "think_token": "<|think|>",
+  "tokenizer_class": "GemmaTokenizer",
+  "unk_token": "<unk>"
+}

26_09_08_sft_email_v4_v3/export/gemma4/210/chat_template.jinja

@@ -0,0 +1,390 @@
+{#
+  Template: Google Gemma 4 Canonical Chat Template
+  Author: Google Gemma Engineering Team
+  Published: 2026-07-09
+  Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
+#}
+{%- macro format_parameters(properties, required, filter_keys=false) -%}
+    {%- set standard_keys = ['properties', 'required', 'nullable'] -%}
+    {%- set ns = namespace(found_first=false) -%}
+    {%- for key, value in properties | dictsort -%}
+        {%- set add_comma = false -%}
+        {%- if not filter_keys or key not in standard_keys -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {{ key }}:{
+            {%- if value['description'] -%}
+                description:<|"|>{{ value['description'] }}<|"|>
+                {%- set add_comma = true -%}
+            {%- endif -%}
+            {%- if value['type'] | upper == 'STRING' -%}
+                {%- if value['enum'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    enum:{{ format_argument(value['enum']) }}
+                {%- endif -%}
+            {%- elif value['type'] | upper == 'ARRAY' -%}
+                {%- if value['items'] is mapping and value['items'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    items:{
+                    {%- set ns_items = namespace(found_first=false) -%}
+                    {%- for item_key, item_value in value['items'] | dictsort -%}
+                        {%- if item_value is not none -%}
+                            {%- if ns_items.found_first %},{% endif -%}
+                            {%- set ns_items.found_first = true -%}
+                            {%- if item_key == 'properties' -%}
+                                properties:{
+                                {%- if item_value is mapping -%}
+                                    {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
+                                {%- endif -%}
+                                }
+                            {%- elif item_key == 'required' -%}
+                                required:[
+                                {%- for req_item in item_value -%}
+                                    <|"|>{{- req_item -}}<|"|>
+                                    {%- if not loop.last %},{% endif -%}
+                                {%- endfor -%}
+                                ]
+                            {%- elif item_key == 'type' -%}
+                                {%- if item_value is string -%}
+                                    type:{{ format_argument(item_value | upper) }}
+                                {%- else -%}
+                                    type:{{ format_argument(item_value | map('upper') | list) }}
+                                {%- endif -%}
+                            {%- else -%}
+                                {{ item_key }}:{{ format_argument(item_value) }}
+                            {%- endif -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                    }
+                {%- endif -%}
+            {%- endif -%}
+            {%- if value['nullable'] %}
+                {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                nullable:true
+            {%- endif -%}
+            {%- if value['type'] | upper == 'OBJECT' -%}
+                {%- if value['properties'] is defined and value['properties'] is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value['properties'], value['required'] | default([])) -}}
+                    }
+                {%- elif value is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}
+                    }
+                {%- endif -%}
+                {%- if value['required'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    required:[
+                    {%- for item in value['required'] | default([]) -%}
+                        <|"|>{{- item -}}<|"|>
+                        {%- if not loop.last %},{% endif -%}
+                    {%- endfor -%}
+                    ]
+                {%- endif -%}
+            {%- endif -%}
+            {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+            type:<|"|>{{ value['type'] | upper }}<|"|>}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endmacro -%}
+{%- macro format_function_declaration(tool_data) -%}
+    declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
+    {%- set params = tool_data['function']['parameters'] -%}
+    {%- if params -%}
+        ,parameters:{
+        {%- if params['properties'] -%}
+            properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
+        {%- endif -%}
+        {%- if params['required'] -%}
+            required:[
+            {%- for item in params['required'] -%}
+                <|"|>{{- item -}}<|"|>
+                {{- ',' if not loop.last -}}
+            {%- endfor -%}
+            ],
+        {%- endif -%}
+        {%- if params['type'] -%}
+            type:<|"|>{{- params['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    {%- if 'response' in tool_data['function'] -%}
+        {%- set response_declaration = tool_data['function']['response'] -%}
+        ,response:{
+        {%- if response_declaration['description'] -%}
+            description:<|"|>{{- response_declaration['description'] -}}<|"|>,
+        {%- endif -%}
+        {%- if response_declaration['type'] | upper == 'OBJECT' -%}
+            type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    }
+{%- endmacro -%}
+{%- macro format_argument(argument, escape_keys=True) -%}
+    {%- if argument is none -%}
+        {{- 'null' -}}
+    {%- elif argument is string -%}
+        {{- '<|"|>' + argument + '<|"|>' -}}
+    {%- elif argument is boolean -%}
+        {{- 'true' if argument else 'false' -}}
+    {%- elif argument is mapping -%}
+        {{- '{' -}}
+        {%- set ns = namespace(found_first=false) -%}
+        {%- for key, value in argument | dictsort -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {%- if escape_keys -%}
+                {{- '<|"|>' + key + '<|"|>' -}}
+            {%- else -%}
+                {{- key -}}
+            {%- endif -%}
+            :{{- format_argument(value, escape_keys=escape_keys) -}}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- elif argument is sequence -%}
+        {{- '[' -}}
+        {%- for item in argument -%}
+            {{- format_argument(item, escape_keys=escape_keys) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- ']' -}}
+    {%- else -%}
+        {{- argument -}}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro strip_thinking(text) -%}
+    {%- set ns = namespace(result='') -%}
+    {%- for part in text.split('<channel|>') -%}
+        {%- if '<|channel>' in part -%}
+            {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
+        {%- else -%}
+            {%- set ns.result = ns.result + part -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {{- ns.result | trim -}}
+{%- endmacro -%}
+
+{%- macro format_tool_response_block(tool_name, response) -%}
+    {{- '<|tool_response>' -}}
+    {%- if response is mapping -%}
+        {{- 'response:' + tool_name + '{' -}}
+        {%- for key, value in response | dictsort -%}
+            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- else -%}
+        {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
+    {%- endif -%}
+    {{- '<tool_response|>' -}}
+{%- endmacro -%}
+
+{#- ===== SETUP ===== -#}
+{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
+{%- set loop_messages = messages -%}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
+{{- bos_token -}}
+{#- Handle System/Tool Definitions Block -#}
+{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
+    {{- '<|turn>system\n' -}}
+    {#- Inject Thinking token at the very top of the FIRST system turn -#}
+    {%- if enable_thinking -%}
+        {{- '<|think|>\n' -}}
+        {%- set ns.prev_message_type = 'think' -%}
+    {%- endif -%}
+    {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
+        {%- if messages[0]['content'] is string -%}
+            {{- messages[0]['content'] | trim -}}
+        {%- elif messages[0]['content'] is sequence -%}
+            {%- for item in messages[0]['content'] -%}
+                {{- item['text'] | trim + ' '-}}
+            {%- endfor -%}
+        {%- endif -%}
+        {%- set loop_messages = messages[1:] -%}
+    {%- endif -%}
+    {%- if tools -%}
+        {%- for tool in tools %}
+            {{- '<|tool>' -}}
+            {{- format_function_declaration(tool) | trim -}}
+            {{- '<tool|>' -}}
+        {%- endfor %}
+        {%- set ns.prev_message_type = 'tool' -%}
+    {%- endif -%}
+    {{- '<turn|>\n' -}}
+{%- endif %}
+
+{#- Pre-scan: find last user message index for reasoning guard -#}
+{%- set ns_turn = namespace(last_user_idx=-1) -%}
+{%- for i in range(loop_messages | length) -%}
+    {%- if loop_messages[i]['role'] == 'user' -%}
+        {%- set ns_turn.last_user_idx = i -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Loop through messages -#}
+{%- for message in loop_messages -%}
+    {%- if message['role'] != 'tool' -%}
+    {%- set ns.prev_message_type = None -%}
+    {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
+    {#- Detect continuation using tracked state — O(1) instead of O(n) backward scan -#}
+    {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
+    {%- if not continue_same_model_turn -%}
+        {{- '<|turn>' + role + '\n' }}
+
+    {%- endif -%}
+
+    {#- Render reasoning/reasoning_content as thinking channel -#}
+    {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
+    {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}
+    {%- if thinking_text and thinking_gate -%}
+        {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
+    {%- endif -%}
+
+            {%- if message.get('tool_calls') -%}
+                {%- for tool_call in message.get('tool_calls') -%}
+                    {%- set function = tool_call['function'] -%}
+                    {{- '<|tool_call>call:' + function['name'] + '{' -}}
+                    {%- if function['arguments'] is mapping -%}
+                        {%- set ns_args = namespace(found_first=false) -%}
+                        {%- for key, value in function['arguments'] | dictsort -%}
+                            {%- if ns_args.found_first %},{% endif -%}
+                            {%- set ns_args.found_first = true -%}
+                            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+                        {%- endfor -%}
+                    {%- elif function['arguments'] is none -%}
+                    {%- else -%}
+                        {{- raise_exception(
+                            "chat_template: tool_calls[].function.arguments must be a "
+                            "JSON object (mapping), not a string. Deserialize arguments "
+                            "before passing to the template."
+                        ) -}}
+                    {%- endif -%}
+                    {{- '}<tool_call|>' -}}
+                {%- endfor -%}
+                {%- set ns.prev_message_type = 'tool_call' -%}
+            {%- endif -%}
+
+            {%- set ns_tr_out = namespace(flag=false) -%}
+            {%- if message.get('tool_responses') -%}
+                {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
+                {%- for tool_response in message.get('tool_responses') -%}
+                    {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
+                    {%- set ns_tr_out.flag = true -%}
+                    {%- set ns.prev_message_type = 'tool_response' -%}
+                {%- endfor -%}
+            {%- elif message.get('tool_calls') -%}
+                {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
+                {%- set ns_tool_scan = namespace(stopped=false) -%}
+                {%- for k in range(loop.index0 + 1, loop_messages | length) -%}
+                    {%- if ns_tool_scan.stopped -%}
+                    {%- elif loop_messages[k]['role'] != 'tool' -%}
+                        {%- set ns_tool_scan.stopped = true -%}
+                    {%- else -%}
+                        {%- set follow = loop_messages[k] -%}
+                        {#- Resolve tool_call_id to function name -#}
+                        {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
+                        {%- for tc in message.get('tool_calls') -%}
+                            {%- if tc.get('id') == follow.get('tool_call_id') -%}
+                                {%- set ns_tname.name = tc['function']['name'] -%}
+                            {%- endif -%}
+                        {%- endfor -%}
+                        {#- Handle content as string or content-parts array -#}
+                        {%- set tool_body = follow.get('content') -%}
+                        {%- if tool_body is string -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- elif tool_body is sequence and tool_body is not string -%}
+                            {%- set ns_txt = namespace(s='') -%}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') == 'text' -%}
+                                    {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
+                                {%- endif -%}
+                            {%- endfor -%}
+                            {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') in ['image', 'image_url'] -%}
+                                    {{- '<|image|>' -}}
+                                {%- elif part.get('type') in ['audio', 'input_audio'] -%}
+                                    {{- '<|audio|>' -}}
+                                {%- elif part.get('type') == 'video' -%}
+                                    {{- '<|video|>' -}}
+                                {%- endif -%}
+                            {%- endfor -%}
+                        {%- else -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- endif -%}
+                        {%- set ns_tr_out.flag = true -%}
+                        {%- set ns.prev_message_type = 'tool_response' -%}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+
+            {%- set captured_content -%}
+            {%- if message.get('content') is string -%}
+                {%- if role == 'model' -%}
+                    {{- strip_thinking(message['content']) -}}
+                {%- else -%}
+                    {{- message['content'] | trim -}}
+                {%- endif -%}
+            {%- elif message.get('content') is sequence -%}
+                {%- for item in message['content'] -%}
+                    {%- if item.get('type') == 'text' -%}
+                        {%- if role == 'model' -%}
+                            {{- strip_thinking(item['text']) -}}
+                        {%- else -%}
+                            {{- item['text'] | trim -}}
+                        {%- endif -%}
+                    {%- elif item.get('type') in ['image', 'image_url'] -%}
+                        {{- '<|image|>' -}}
+                    {%- elif item.get('type') in ['audio', 'input_audio'] -%}
+                        {{- '<|audio|>' -}}
+                    {%- elif item.get('type') == 'video' -%}
+                        {{- '<|video|>' -}}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+            {%- endset -%}
+
+            {{- captured_content -}}
+            {%- set has_content = captured_content | trim | length > 0 -%}
+
+        {#- Forward-scan: find next non-tool message role for continuation detection -#}
+        {%- set next_nt = namespace(role=None, found=false) -%}
+        {%- for j in range(loop.index0 + 1, loop_messages | length) -%}
+            {%- if not next_nt.found -%}
+                {%- if loop_messages[j]['role'] != 'tool' -%}
+                    {%- set next_nt.role = loop_messages[j]['role'] -%}
+                    {%- set next_nt.found = true -%}
+                {%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+
+        {%- set continues_into_next = (
+            role == 'model'
+            and next_nt.role == 'assistant'
+            and (not message.get('tool_calls') or ns_tr_out.flag)
+        ) -%}
+
+        {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
+            {{- '<|tool_response>' -}}
+        {%- elif continues_into_next -%}
+        {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}
+            {{- '<turn|>\n' -}}
+        {%- endif -%}
+
+    {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
+    {%- set ns.prev_non_tool_role = message['role'] -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if add_generation_prompt -%}
+    {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
+        {{- '<|turn>model\n' -}}
+        {%- if not enable_thinking -%}
+            {{- '<|channel>thought\n<channel|>' -}}
+        {%- endif -%}
+    {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
+        {{- '<|channel>thought\n' -}}
+    {%- endif -%}
+{%- endif -%}

26_09_08_sft_email_v4_v3/export/gemma4/210/config.json

@@ -0,0 +1,83 @@
+{
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attention_k_eq_v": true,
+  "bos_token_id": 2,
+  "dtype": "bfloat16",
+  "enable_moe_block": true,
+  "eos_token_id": 1,
+  "final_logit_softcapping": 30.0,
+  "global_head_dim": 512,
+  "head_dim": 256,
+  "hidden_activation": "gelu_pytorch_tanh",
+  "hidden_size": 2816,
+  "hidden_size_per_layer_input": 0,
+  "initializer_range": 0.02,
+  "intermediate_size": 2112,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 262144,
+  "model_type": "gemma4_text",
+  "moe_intermediate_size": 704,
+  "num_attention_heads": 16,
+  "num_experts": 128,
+  "num_global_key_value_heads": 2,
+  "num_hidden_layers": 30,
+  "num_key_value_heads": 8,
+  "num_kv_shared_layers": 0,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "full_attention": {
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 1000000.0,
+      "rope_type": "proportional"
+    },
+    "sliding_attention": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    }
+  },
+  "sliding_window": 1024,
+  "tie_word_embeddings": true,
+  "top_k_experts": 8,
+  "use_bidirectional_attention": "vision",
+  "use_cache": true,
+  "use_double_wide_mlp": false,
+  "vocab_size": 262144,
+  "vocab_size_per_layer_input": 262144,
+  "architectures": [
+    "Gemma4ForCausalLM"
+  ],
+  "transformers_version": "5.5.0.dev0"
+}

26_09_08_sft_email_v4_v3/export/gemma4/210/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 2,
+  "do_sample": true,
+  "eos_token_id": [
+    1,
+    106,
+    50
+  ],
+  "pad_token_id": 0,
+  "temperature": 1.0,
+  "top_k": 64,
+  "top_p": 0.95,
+  "transformers_version": "5.5.0.dev0"
+}

26_09_08_sft_email_v4_v3/export/gemma4/210/model.safetensors.index.json

@@ -0,0 +1,1053 @@
+{
+    "metadata": {
+        "fast_llm_metadata": {
+            "fast_llm_version": "0.4.0",
+            "model": "gpt",
+            "format": "gemma4",
+            "config": {
+                "type": "gpt",
+                "base_model": {
+                    "decoder": {
+                        "type": "pattern",
+                        "blocks": {
+                            "full_attention": {
+                                "type": "decoder",
+                                "mixer": {
+                                    "type": "attention",
+                                    "rotary": {
+                                        "type": "proportional",
+                                        "theta": 1000000.0,
+                                        "partial_rotary_factor": 0.25
+                                    },
+                                    "heads": 16,
+                                    "head_groups": 2,
+                                    "head_size": 512,
+                                    "add_linear_biases": false,
+                                    "dropout": 0.0,
+                                    "softmax_scale_power": 0.0,
+                                    "implementation": "sdpa",
+                                    "query_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "key_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "value_norm": {
+                                        "type": "fixed_rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "shared_key_value": true
+                                },
+                                "mlp": {
+                                    "type": "hybrid_moe",
+                                    "dense": {
+                                        "type": "mlp",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 2112,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full"
+                                    },
+                                    "routed": {
+                                        "type": "moe",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 704,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full",
+                                        "router_normalization": {
+                                            "type": "fixed_rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "router_scale": {
+                                            "enabled": true
+                                        },
+                                        "router_input_scale": 0.018844459036110227,
+                                        "router_per_expert_scale": {
+                                            "enabled": true
+                                        },
+                                        "experts": 128,
+                                        "experts_per_token": 8,
+                                        "auxiliary_loss_coefficient": 0.0
+                                    }
+                                },
+                                "normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "pre_mlp_normalization": {
+                                    "type": "none"
+                                },
+                                "post_mixer_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "post_mlp_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "output_scale": {
+                                    "lr_scale": 0.0,
+                                    "enabled": true
+                                }
+                            },
+                            "sliding_attention": {
+                                "type": "decoder",
+                                "mixer": {
+                                    "type": "attention",
+                                    "rotary": {
+                                        "type": "default",
+                                        "theta": 10000.0
+                                    },
+                                    "heads": 16,
+                                    "head_groups": 8,
+                                    "head_size": 256,
+                                    "add_linear_biases": false,
+                                    "dropout": 0.0,
+                                    "window_size": 1024,
+                                    "softmax_scale_power": 0.0,
+                                    "implementation": "flash",
+                                    "query_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "key_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "value_norm": {
+                                        "type": "fixed_rms_norm",
+                                        "epsilon": 1e-06
+                                    }
+                                },
+                                "mlp": {
+                                    "type": "hybrid_moe",
+                                    "dense": {
+                                        "type": "mlp",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 2112,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full"
+                                    },
+                                    "routed": {
+                                        "type": "moe",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 704,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full",
+                                        "router_normalization": {
+                                            "type": "fixed_rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "router_scale": {
+                                            "enabled": true
+                                        },
+                                        "router_input_scale": 0.018844459036110227,
+                                        "router_per_expert_scale": {
+                                            "enabled": true
+                                        },
+                                        "experts": 128,
+                                        "experts_per_token": 8,
+                                        "auxiliary_loss_coefficient": 0.0
+                                    }
+                                },
+                                "normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "pre_mlp_normalization": {
+                                    "type": "none"
+                                },
+                                "post_mixer_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "post_mlp_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "output_scale": {
+                                    "lr_scale": 0.0,
+                                    "enabled": true
+                                }
+                            }
+                        },
+                        "pattern": [
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention"
+                        ],
+                        "num_blocks": 30
+                    },
+                    "embeddings": {
+                        "vocab_size": 262144,
+                        "vocab_parallel": false,
+                        "embedding_scale": 53.0659966456864
+                    },
+                    "head": {
+                        "normalization": {
+                            "type": "rms_norm",
+                            "epsilon": 1e-06
+                        },
+                        "cross_entropy_splits": 8,
+                        "final_logit_softcap": 30.0
+                    },
+                    "tied_embedding_weight": true,
+                    "hidden_size": 2816,
+                    "peft": {
+                        "type": "none"
+                    }
+                },
+                "multi_stage": {
+                    "zero_stage": 3
+                },
+                "distributed": {
+                    "tensor_parallel": 2,
+                    "sequence_tensor_parallel": true,
+                    "sequence_data_parallel": 4,
+                    "world_size": 32,
+                    "rank": 0,
+                    "local_world_size": 8,
+                    "timeout": 3600.0,
+                    "compute_dtype": "bfloat16"
+                }
+            },
+            "shards": [
+                "weights"
+            ],
+            "metadata": {
+                "optimizer": {
+                    "current_step": 210,
+                    "grad_scaler": {
+                        "type": "NoopGradScaler"
+                    }
+                },
+                "completed_steps": 210,
+                "metrics": {
+                    "training": {
+                        "batch_size": 2097152,
+                        "load_balancing_loss": 815.46337890625,
+                        "lm_head_loss": 0.11222395300865173,
+                        "total_steps": 210,
+                        "completed_steps": 210,
+                        "consumed_tokens": 440401920,
+                        "percent_done": 100.0,
+                        "step_time_ms": 28241.050300188363,
+                        "step_time_average_ms": 30303.581654688434,
+                        "remaining_time": 0.0,
+                        "completion_time": 1788873090.3226085,
+                        "skipped_iters": 0,
+                        "nan_iters": 0,
+                        "model_tflops": 59447.884978233706,
+                        "hardware_tflops": 73753.52245463323,
+                        "tokens_per_sec_per_gpu": 2320.5935793245935,
+                        "run": 0,
+                        "grad_norm": 2.482532262802124,
+                        "learning_rate": 0.0,
+                        "loss_scale": 1.0,
+                        "reserved": 55562.0,
+                        "allocated": 21478.49755859375,
+                        "max_allocated": 54800.1123046875,
+                        "max_reserved": 55562.0,
+                        "global_max_reserved": 74782.0
+                    }
+                }
+            }
+        },
+        "model_config": {
+            "vocab_size": 262144,
+            "num_attention_heads": 16,
+            "num_key_value_heads": 8,
+            "head_dim": 256,
+            "global_head_dim": 512,
+            "num_global_key_value_heads": 2,
+            "attention_bias": false,
+            "attention_dropout": 0.0,
+            "sliding_window": 1024,
+            "rms_norm_eps": 1e-06,
+            "attention_k_eq_v": true,
+            "rope_parameters": {
+                "sliding_attention": {
+                    "rope_type": "default",
+                    "rope_theta": 10000.0
+                },
+                "full_attention": {
+                    "rope_type": "proportional",
+                    "rope_theta": 1000000.0,
+                    "partial_rotary_factor": 0.25
+                }
+            },
+            "intermediate_size": 2112,
+            "hidden_activation": "gelu_pytorch_tanh",
+            "num_experts": 128,
+            "top_k_experts": 8,
+            "moe_intermediate_size": 704,
+            "enable_moe_block": true,
+            "num_hidden_layers": 30,
+            "layer_types": [
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention"
+            ],
+            "final_logit_softcapping": 30.0,
+            "tie_word_embeddings": true,
+            "hidden_size": 2816,
+            "hidden_size_per_layer_input": 0,
+            "use_bidirectional_attention": null,
+            "model_type": "gemma4_text",
+            "architectures": [
+                "Gemma4ForCausalLM"
+            ]
+        },
+        "format": "pt"
+    },
+    "weight_map": {
+        "model.embed_tokens.weight": "model_0.safetensors",
+        "model.layers.0.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.0.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.0.router.scale": "model_0.safetensors",
+        "model.layers.0.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.0.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.0.router.proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.0.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.0.experts.down_proj": "model_0.safetensors",
+        "model.layers.0.layer_scalar": "model_0.safetensors",
+        "model.layers.1.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.1.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.1.router.scale": "model_0.safetensors",
+        "model.layers.1.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.1.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.1.router.proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.1.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.1.experts.down_proj": "model_0.safetensors",
+        "model.layers.1.layer_scalar": "model_0.safetensors",
+        "model.layers.2.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.2.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.2.router.scale": "model_0.safetensors",
+        "model.layers.2.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.2.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.2.router.proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.2.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.2.experts.down_proj": "model_0.safetensors",
+        "model.layers.2.layer_scalar": "model_0.safetensors",
+        "model.layers.3.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.3.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.3.router.scale": "model_0.safetensors",
+        "model.layers.3.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.3.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.3.router.proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.3.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.3.experts.down_proj": "model_0.safetensors",
+        "model.layers.3.layer_scalar": "model_0.safetensors",
+        "model.layers.4.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.4.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.4.router.scale": "model_0.safetensors",
+        "model.layers.4.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.4.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.4.router.proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.4.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.4.experts.down_proj": "model_1.safetensors",
+        "model.layers.4.layer_scalar": "model_1.safetensors",
+        "model.layers.5.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.5.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.5.router.scale": "model_1.safetensors",
+        "model.layers.5.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.5.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.5.router.proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.5.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.5.experts.down_proj": "model_1.safetensors",
+        "model.layers.5.layer_scalar": "model_1.safetensors",
+        "model.layers.6.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.6.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.6.router.scale": "model_1.safetensors",
+        "model.layers.6.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.6.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.6.router.proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.6.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.6.experts.down_proj": "model_1.safetensors",
+        "model.layers.6.layer_scalar": "model_1.safetensors",
+        "model.layers.7.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.7.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.7.router.scale": "model_1.safetensors",
+        "model.layers.7.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.7.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.7.router.proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.7.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.7.experts.down_proj": "model_1.safetensors",
+        "model.layers.7.layer_scalar": "model_1.safetensors",
+        "model.layers.8.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.8.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.8.router.scale": "model_1.safetensors",
+        "model.layers.8.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.8.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.8.router.proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.8.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.8.experts.down_proj": "model_1.safetensors",
+        "model.layers.8.layer_scalar": "model_1.safetensors",
+        "model.layers.9.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.9.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.9.router.scale": "model_1.safetensors",
+        "model.layers.9.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.9.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.9.router.proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.9.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.9.experts.down_proj": "model_1.safetensors",
+        "model.layers.9.layer_scalar": "model_2.safetensors",
+        "model.layers.10.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.10.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.10.router.scale": "model_2.safetensors",
+        "model.layers.10.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.10.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.10.router.proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.10.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.10.experts.down_proj": "model_2.safetensors",
+        "model.layers.10.layer_scalar": "model_2.safetensors",
+        "model.layers.11.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.11.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.11.router.scale": "model_2.safetensors",
+        "model.layers.11.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.11.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.11.router.proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.11.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.11.experts.down_proj": "model_2.safetensors",
+        "model.layers.11.layer_scalar": "model_2.safetensors",
+        "model.layers.12.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.12.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.12.router.scale": "model_2.safetensors",
+        "model.layers.12.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.12.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.12.router.proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.12.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.12.experts.down_proj": "model_2.safetensors",
+        "model.layers.12.layer_scalar": "model_2.safetensors",
+        "model.layers.13.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.13.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.13.router.scale": "model_2.safetensors",
+        "model.layers.13.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.13.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.13.router.proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.13.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.13.experts.down_proj": "model_2.safetensors",
+        "model.layers.13.layer_scalar": "model_2.safetensors",
+        "model.layers.14.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.14.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.14.router.scale": "model_2.safetensors",
+        "model.layers.14.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.14.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.14.router.proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.14.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.14.experts.down_proj": "model_2.safetensors",
+        "model.layers.14.layer_scalar": "model_2.safetensors",
+        "model.layers.15.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.15.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.15.router.scale": "model_2.safetensors",
+        "model.layers.15.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.15.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.15.router.proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.15.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.15.experts.down_proj": "model_3.safetensors",
+        "model.layers.15.layer_scalar": "model_3.safetensors",
+        "model.layers.16.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.16.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.16.router.scale": "model_3.safetensors",
+        "model.layers.16.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.16.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.16.router.proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.16.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.16.experts.down_proj": "model_3.safetensors",
+        "model.layers.16.layer_scalar": "model_3.safetensors",
+        "model.layers.17.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.17.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.17.router.scale": "model_3.safetensors",
+        "model.layers.17.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.17.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.17.router.proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.17.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.17.experts.down_proj": "model_3.safetensors",
+        "model.layers.17.layer_scalar": "model_3.safetensors",
+        "model.layers.18.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.18.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.18.router.scale": "model_3.safetensors",
+        "model.layers.18.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.18.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.18.router.proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.18.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.18.experts.down_proj": "model_3.safetensors",
+        "model.layers.18.layer_scalar": "model_3.safetensors",
+        "model.layers.19.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.19.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.19.router.scale": "model_3.safetensors",
+        "model.layers.19.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.19.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.19.router.proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.19.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.19.experts.down_proj": "model_3.safetensors",
+        "model.layers.19.layer_scalar": "model_3.safetensors",
+        "model.layers.20.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.20.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.20.router.scale": "model_3.safetensors",
+        "model.layers.20.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.20.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.20.router.proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.20.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.20.experts.down_proj": "model_3.safetensors",
+        "model.layers.20.layer_scalar": "model_4.safetensors",
+        "model.layers.21.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.21.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.21.router.scale": "model_4.safetensors",
+        "model.layers.21.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.21.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.21.router.proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.21.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.21.experts.down_proj": "model_4.safetensors",
+        "model.layers.21.layer_scalar": "model_4.safetensors",
+        "model.layers.22.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.22.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.22.router.scale": "model_4.safetensors",
+        "model.layers.22.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.22.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.22.router.proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.22.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.22.experts.down_proj": "model_4.safetensors",
+        "model.layers.22.layer_scalar": "model_4.safetensors",
+        "model.layers.23.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.23.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.23.router.scale": "model_4.safetensors",
+        "model.layers.23.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.23.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.23.router.proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.23.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.23.experts.down_proj": "model_4.safetensors",
+        "model.layers.23.layer_scalar": "model_4.safetensors",
+        "model.layers.24.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.24.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.24.router.scale": "model_4.safetensors",
+        "model.layers.24.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.24.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.24.router.proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.24.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.24.experts.down_proj": "model_4.safetensors",
+        "model.layers.24.layer_scalar": "model_4.safetensors",
+        "model.layers.25.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.25.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.25.router.scale": "model_4.safetensors",
+        "model.layers.25.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.25.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.25.router.proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.25.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.25.experts.down_proj": "model_4.safetensors",
+        "model.layers.25.layer_scalar": "model_4.safetensors",
+        "model.layers.26.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.26.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.26.router.scale": "model_4.safetensors",
+        "model.layers.26.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.26.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.26.router.proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.26.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.26.experts.down_proj": "model_5.safetensors",
+        "model.layers.26.layer_scalar": "model_5.safetensors",
+        "model.layers.27.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.27.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.27.router.scale": "model_5.safetensors",
+        "model.layers.27.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.27.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.27.router.proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.v_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.27.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.27.experts.down_proj": "model_5.safetensors",
+        "model.layers.27.layer_scalar": "model_5.safetensors",
+        "model.layers.28.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.28.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.28.router.scale": "model_5.safetensors",
+        "model.layers.28.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.28.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.28.router.proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.v_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.28.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.28.experts.down_proj": "model_5.safetensors",
+        "model.layers.28.layer_scalar": "model_5.safetensors",
+        "model.layers.29.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.29.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.29.router.scale": "model_5.safetensors",
+        "model.layers.29.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.29.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.29.router.proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.29.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.29.experts.down_proj": "model_5.safetensors",
+        "model.layers.29.layer_scalar": "model_5.safetensors",
+        "model.norm.weight": "model_5.safetensors"
+    }
+}

26_09_08_sft_email_v4_v3/export/gemma4/210/model_0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:235b37f041fa036c111cda5de1c386e4acd1c02c72ec049d14cd8ec9a51ed8c8
+size 18219689400

26_09_08_sft_email_v4_v3/export/gemma4/210/model_1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53988b57f3f6885db1404438af6643c77807f3ab246fd8f5c103c21680780433
+size 17354616348

26_09_08_sft_email_v4_v3/export/gemma4/210/model_2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d409016595f4d1672722b49f995b09f73292e334c58d715980dbd34d9cc3c047
+size 18580955536

26_09_08_sft_email_v4_v3/export/gemma4/210/model_3.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:583d9bbfd8fc19ea1936117ae1a651e6469205767e71b6a4a65b8d65eedef1f1
+size 17354616460

26_09_08_sft_email_v4_v3/export/gemma4/210/model_4.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b686f62b6a8ff945889290ab68dfcfdfd7156822fc5fd86f32c79c75f3f28789
+size 18580955512

26_09_08_sft_email_v4_v3/export/gemma4/210/model_5.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdb49e2609c628513f93b5f57ef56edcd487b6256f66a902509231b24da3fdfd
+size 10841862960

26_09_08_sft_email_v4_v3/export/gemma4/210/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f
+size 32169626

26_09_08_sft_email_v4_v3/export/gemma4/210/tokenizer_config.json

@@ -0,0 +1,74 @@
+{
+  "audio_token": "<|audio|>",
+  "backend": "tokenizers",
+  "boa_token": "<|audio>",
+  "boi_token": "<|image>",
+  "bos_token": "<bos>",
+  "eoa_token": "<audio|>",
+  "eoc_token": "<channel|>",
+  "eoi_token": "<image|>",
+  "eos_token": "<eos>",
+  "eot_token": "<turn|>",
+  "escape_token": "<|\"|>",
+  "etc_token": "<tool_call|>",
+  "etd_token": "<tool|>",
+  "etr_token": "<tool_response|>",
+  "extra_special_tokens": [
+    "<|video|>"
+  ],
+  "image_token": "<|image|>",
+  "mask_token": "<mask>",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "padding_side": "left",
+  "processor_class": "Gemma4Processor",
+  "response_schema": {
+    "type": "object",
+    "properties": {
+      "role": {
+        "const": "assistant"
+      },
+      "thinking": {
+        "type": "string"
+      },
+      "content": {
+        "type": "string"
+      },
+      "tool_calls": {
+        "x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>",
+        "type": "array",
+        "items": {
+          "type": "object",
+          "properties": {
+            "type": {
+              "const": "function"
+            },
+            "function": {
+              "type": "object",
+              "x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})",
+              "properties": {
+                "name": {
+                  "type": "string"
+                },
+                "arguments": {
+                  "type": "object",
+                  "x-parser": "gemma4-tool-call",
+                  "additionalProperties": {}
+                }
+              }
+            }
+          }
+        }
+      }
+    },
+    "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
+  },
+  "soc_token": "<|channel>",
+  "sot_token": "<|turn>",
+  "stc_token": "<|tool_call>",
+  "std_token": "<|tool>",
+  "str_token": "<|tool_response>",
+  "think_token": "<|think|>",
+  "tokenizer_class": "GemmaTokenizer",
+  "unk_token": "<unk>"
+}

26_09_08_sft_email_v4_v3/export/gemma4/70/chat_template.jinja

@@ -0,0 +1,390 @@
+{#
+  Template: Google Gemma 4 Canonical Chat Template
+  Author: Google Gemma Engineering Team
+  Published: 2026-07-09
+  Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
+#}
+{%- macro format_parameters(properties, required, filter_keys=false) -%}
+    {%- set standard_keys = ['properties', 'required', 'nullable'] -%}
+    {%- set ns = namespace(found_first=false) -%}
+    {%- for key, value in properties | dictsort -%}
+        {%- set add_comma = false -%}
+        {%- if not filter_keys or key not in standard_keys -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {{ key }}:{
+            {%- if value['description'] -%}
+                description:<|"|>{{ value['description'] }}<|"|>
+                {%- set add_comma = true -%}
+            {%- endif -%}
+            {%- if value['type'] | upper == 'STRING' -%}
+                {%- if value['enum'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    enum:{{ format_argument(value['enum']) }}
+                {%- endif -%}
+            {%- elif value['type'] | upper == 'ARRAY' -%}
+                {%- if value['items'] is mapping and value['items'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    items:{
+                    {%- set ns_items = namespace(found_first=false) -%}
+                    {%- for item_key, item_value in value['items'] | dictsort -%}
+                        {%- if item_value is not none -%}
+                            {%- if ns_items.found_first %},{% endif -%}
+                            {%- set ns_items.found_first = true -%}
+                            {%- if item_key == 'properties' -%}
+                                properties:{
+                                {%- if item_value is mapping -%}
+                                    {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
+                                {%- endif -%}
+                                }
+                            {%- elif item_key == 'required' -%}
+                                required:[
+                                {%- for req_item in item_value -%}
+                                    <|"|>{{- req_item -}}<|"|>
+                                    {%- if not loop.last %},{% endif -%}
+                                {%- endfor -%}
+                                ]
+                            {%- elif item_key == 'type' -%}
+                                {%- if item_value is string -%}
+                                    type:{{ format_argument(item_value | upper) }}
+                                {%- else -%}
+                                    type:{{ format_argument(item_value | map('upper') | list) }}
+                                {%- endif -%}
+                            {%- else -%}
+                                {{ item_key }}:{{ format_argument(item_value) }}
+                            {%- endif -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                    }
+                {%- endif -%}
+            {%- endif -%}
+            {%- if value['nullable'] %}
+                {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                nullable:true
+            {%- endif -%}
+            {%- if value['type'] | upper == 'OBJECT' -%}
+                {%- if value['properties'] is defined and value['properties'] is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value['properties'], value['required'] | default([])) -}}
+                    }
+                {%- elif value is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}
+                    }
+                {%- endif -%}
+                {%- if value['required'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    required:[
+                    {%- for item in value['required'] | default([]) -%}
+                        <|"|>{{- item -}}<|"|>
+                        {%- if not loop.last %},{% endif -%}
+                    {%- endfor -%}
+                    ]
+                {%- endif -%}
+            {%- endif -%}
+            {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+            type:<|"|>{{ value['type'] | upper }}<|"|>}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endmacro -%}
+{%- macro format_function_declaration(tool_data) -%}
+    declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
+    {%- set params = tool_data['function']['parameters'] -%}
+    {%- if params -%}
+        ,parameters:{
+        {%- if params['properties'] -%}
+            properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
+        {%- endif -%}
+        {%- if params['required'] -%}
+            required:[
+            {%- for item in params['required'] -%}
+                <|"|>{{- item -}}<|"|>
+                {{- ',' if not loop.last -}}
+            {%- endfor -%}
+            ],
+        {%- endif -%}
+        {%- if params['type'] -%}
+            type:<|"|>{{- params['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    {%- if 'response' in tool_data['function'] -%}
+        {%- set response_declaration = tool_data['function']['response'] -%}
+        ,response:{
+        {%- if response_declaration['description'] -%}
+            description:<|"|>{{- response_declaration['description'] -}}<|"|>,
+        {%- endif -%}
+        {%- if response_declaration['type'] | upper == 'OBJECT' -%}
+            type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    }
+{%- endmacro -%}
+{%- macro format_argument(argument, escape_keys=True) -%}
+    {%- if argument is none -%}
+        {{- 'null' -}}
+    {%- elif argument is string -%}
+        {{- '<|"|>' + argument + '<|"|>' -}}
+    {%- elif argument is boolean -%}
+        {{- 'true' if argument else 'false' -}}
+    {%- elif argument is mapping -%}
+        {{- '{' -}}
+        {%- set ns = namespace(found_first=false) -%}
+        {%- for key, value in argument | dictsort -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {%- if escape_keys -%}
+                {{- '<|"|>' + key + '<|"|>' -}}
+            {%- else -%}
+                {{- key -}}
+            {%- endif -%}
+            :{{- format_argument(value, escape_keys=escape_keys) -}}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- elif argument is sequence -%}
+        {{- '[' -}}
+        {%- for item in argument -%}
+            {{- format_argument(item, escape_keys=escape_keys) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- ']' -}}
+    {%- else -%}
+        {{- argument -}}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro strip_thinking(text) -%}
+    {%- set ns = namespace(result='') -%}
+    {%- for part in text.split('<channel|>') -%}
+        {%- if '<|channel>' in part -%}
+            {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
+        {%- else -%}
+            {%- set ns.result = ns.result + part -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {{- ns.result | trim -}}
+{%- endmacro -%}
+
+{%- macro format_tool_response_block(tool_name, response) -%}
+    {{- '<|tool_response>' -}}
+    {%- if response is mapping -%}
+        {{- 'response:' + tool_name + '{' -}}
+        {%- for key, value in response | dictsort -%}
+            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- else -%}
+        {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
+    {%- endif -%}
+    {{- '<tool_response|>' -}}
+{%- endmacro -%}
+
+{#- ===== SETUP ===== -#}
+{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
+{%- set loop_messages = messages -%}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
+{{- bos_token -}}
+{#- Handle System/Tool Definitions Block -#}
+{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
+    {{- '<|turn>system\n' -}}
+    {#- Inject Thinking token at the very top of the FIRST system turn -#}
+    {%- if enable_thinking -%}
+        {{- '<|think|>\n' -}}
+        {%- set ns.prev_message_type = 'think' -%}
+    {%- endif -%}
+    {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
+        {%- if messages[0]['content'] is string -%}
+            {{- messages[0]['content'] | trim -}}
+        {%- elif messages[0]['content'] is sequence -%}
+            {%- for item in messages[0]['content'] -%}
+                {{- item['text'] | trim + ' '-}}
+            {%- endfor -%}
+        {%- endif -%}
+        {%- set loop_messages = messages[1:] -%}
+    {%- endif -%}
+    {%- if tools -%}
+        {%- for tool in tools %}
+            {{- '<|tool>' -}}
+            {{- format_function_declaration(tool) | trim -}}
+            {{- '<tool|>' -}}
+        {%- endfor %}
+        {%- set ns.prev_message_type = 'tool' -%}
+    {%- endif -%}
+    {{- '<turn|>\n' -}}
+{%- endif %}
+
+{#- Pre-scan: find last user message index for reasoning guard -#}
+{%- set ns_turn = namespace(last_user_idx=-1) -%}
+{%- for i in range(loop_messages | length) -%}
+    {%- if loop_messages[i]['role'] == 'user' -%}
+        {%- set ns_turn.last_user_idx = i -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Loop through messages -#}
+{%- for message in loop_messages -%}
+    {%- if message['role'] != 'tool' -%}
+    {%- set ns.prev_message_type = None -%}
+    {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
+    {#- Detect continuation using tracked state — O(1) instead of O(n) backward scan -#}
+    {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
+    {%- if not continue_same_model_turn -%}
+        {{- '<|turn>' + role + '\n' }}
+
+    {%- endif -%}
+
+    {#- Render reasoning/reasoning_content as thinking channel -#}
+    {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
+    {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}
+    {%- if thinking_text and thinking_gate -%}
+        {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
+    {%- endif -%}
+
+            {%- if message.get('tool_calls') -%}
+                {%- for tool_call in message.get('tool_calls') -%}
+                    {%- set function = tool_call['function'] -%}
+                    {{- '<|tool_call>call:' + function['name'] + '{' -}}
+                    {%- if function['arguments'] is mapping -%}
+                        {%- set ns_args = namespace(found_first=false) -%}
+                        {%- for key, value in function['arguments'] | dictsort -%}
+                            {%- if ns_args.found_first %},{% endif -%}
+                            {%- set ns_args.found_first = true -%}
+                            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+                        {%- endfor -%}
+                    {%- elif function['arguments'] is none -%}
+                    {%- else -%}
+                        {{- raise_exception(
+                            "chat_template: tool_calls[].function.arguments must be a "
+                            "JSON object (mapping), not a string. Deserialize arguments "
+                            "before passing to the template."
+                        ) -}}
+                    {%- endif -%}
+                    {{- '}<tool_call|>' -}}
+                {%- endfor -%}
+                {%- set ns.prev_message_type = 'tool_call' -%}
+            {%- endif -%}
+
+            {%- set ns_tr_out = namespace(flag=false) -%}
+            {%- if message.get('tool_responses') -%}
+                {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
+                {%- for tool_response in message.get('tool_responses') -%}
+                    {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
+                    {%- set ns_tr_out.flag = true -%}
+                    {%- set ns.prev_message_type = 'tool_response' -%}
+                {%- endfor -%}
+            {%- elif message.get('tool_calls') -%}
+                {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
+                {%- set ns_tool_scan = namespace(stopped=false) -%}
+                {%- for k in range(loop.index0 + 1, loop_messages | length) -%}
+                    {%- if ns_tool_scan.stopped -%}
+                    {%- elif loop_messages[k]['role'] != 'tool' -%}
+                        {%- set ns_tool_scan.stopped = true -%}
+                    {%- else -%}
+                        {%- set follow = loop_messages[k] -%}
+                        {#- Resolve tool_call_id to function name -#}
+                        {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
+                        {%- for tc in message.get('tool_calls') -%}
+                            {%- if tc.get('id') == follow.get('tool_call_id') -%}
+                                {%- set ns_tname.name = tc['function']['name'] -%}
+                            {%- endif -%}
+                        {%- endfor -%}
+                        {#- Handle content as string or content-parts array -#}
+                        {%- set tool_body = follow.get('content') -%}
+                        {%- if tool_body is string -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- elif tool_body is sequence and tool_body is not string -%}
+                            {%- set ns_txt = namespace(s='') -%}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') == 'text' -%}
+                                    {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
+                                {%- endif -%}
+                            {%- endfor -%}
+                            {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') in ['image', 'image_url'] -%}
+                                    {{- '<|image|>' -}}
+                                {%- elif part.get('type') in ['audio', 'input_audio'] -%}
+                                    {{- '<|audio|>' -}}
+                                {%- elif part.get('type') == 'video' -%}
+                                    {{- '<|video|>' -}}
+                                {%- endif -%}
+                            {%- endfor -%}
+                        {%- else -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- endif -%}
+                        {%- set ns_tr_out.flag = true -%}
+                        {%- set ns.prev_message_type = 'tool_response' -%}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+
+            {%- set captured_content -%}
+            {%- if message.get('content') is string -%}
+                {%- if role == 'model' -%}
+                    {{- strip_thinking(message['content']) -}}
+                {%- else -%}
+                    {{- message['content'] | trim -}}
+                {%- endif -%}
+            {%- elif message.get('content') is sequence -%}
+                {%- for item in message['content'] -%}
+                    {%- if item.get('type') == 'text' -%}
+                        {%- if role == 'model' -%}
+                            {{- strip_thinking(item['text']) -}}
+                        {%- else -%}
+                            {{- item['text'] | trim -}}
+                        {%- endif -%}
+                    {%- elif item.get('type') in ['image', 'image_url'] -%}
+                        {{- '<|image|>' -}}
+                    {%- elif item.get('type') in ['audio', 'input_audio'] -%}
+                        {{- '<|audio|>' -}}
+                    {%- elif item.get('type') == 'video' -%}
+                        {{- '<|video|>' -}}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+            {%- endset -%}
+
+            {{- captured_content -}}
+            {%- set has_content = captured_content | trim | length > 0 -%}
+
+        {#- Forward-scan: find next non-tool message role for continuation detection -#}
+        {%- set next_nt = namespace(role=None, found=false) -%}
+        {%- for j in range(loop.index0 + 1, loop_messages | length) -%}
+            {%- if not next_nt.found -%}
+                {%- if loop_messages[j]['role'] != 'tool' -%}
+                    {%- set next_nt.role = loop_messages[j]['role'] -%}
+                    {%- set next_nt.found = true -%}
+                {%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+
+        {%- set continues_into_next = (
+            role == 'model'
+            and next_nt.role == 'assistant'
+            and (not message.get('tool_calls') or ns_tr_out.flag)
+        ) -%}
+
+        {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
+            {{- '<|tool_response>' -}}
+        {%- elif continues_into_next -%}
+        {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}
+            {{- '<turn|>\n' -}}
+        {%- endif -%}
+
+    {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
+    {%- set ns.prev_non_tool_role = message['role'] -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if add_generation_prompt -%}
+    {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
+        {{- '<|turn>model\n' -}}
+        {%- if not enable_thinking -%}
+            {{- '<|channel>thought\n<channel|>' -}}
+        {%- endif -%}
+    {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
+        {{- '<|channel>thought\n' -}}
+    {%- endif -%}
+{%- endif -%}

26_09_08_sft_email_v4_v3/export/gemma4/70/config.json

@@ -0,0 +1,83 @@
+{
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attention_k_eq_v": true,
+  "bos_token_id": 2,
+  "dtype": "bfloat16",
+  "enable_moe_block": true,
+  "eos_token_id": 1,
+  "final_logit_softcapping": 30.0,
+  "global_head_dim": 512,
+  "head_dim": 256,
+  "hidden_activation": "gelu_pytorch_tanh",
+  "hidden_size": 2816,
+  "hidden_size_per_layer_input": 0,
+  "initializer_range": 0.02,
+  "intermediate_size": 2112,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 262144,
+  "model_type": "gemma4_text",
+  "moe_intermediate_size": 704,
+  "num_attention_heads": 16,
+  "num_experts": 128,
+  "num_global_key_value_heads": 2,
+  "num_hidden_layers": 30,
+  "num_key_value_heads": 8,
+  "num_kv_shared_layers": 0,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "full_attention": {
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 1000000.0,
+      "rope_type": "proportional"
+    },
+    "sliding_attention": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    }
+  },
+  "sliding_window": 1024,
+  "tie_word_embeddings": true,
+  "top_k_experts": 8,
+  "use_bidirectional_attention": "vision",
+  "use_cache": true,
+  "use_double_wide_mlp": false,
+  "vocab_size": 262144,
+  "vocab_size_per_layer_input": 262144,
+  "architectures": [
+    "Gemma4ForCausalLM"
+  ],
+  "transformers_version": "5.5.0.dev0"
+}

26_09_08_sft_email_v4_v3/export/gemma4/70/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 2,
+  "do_sample": true,
+  "eos_token_id": [
+    1,
+    106,
+    50
+  ],
+  "pad_token_id": 0,
+  "temperature": 1.0,
+  "top_k": 64,
+  "top_p": 0.95,
+  "transformers_version": "5.5.0.dev0"
+}

26_09_08_sft_email_v4_v3/export/gemma4/70/model.safetensors.index.json

@@ -0,0 +1,1053 @@
+{
+    "metadata": {
+        "fast_llm_metadata": {
+            "fast_llm_version": "0.4.0",
+            "model": "gpt",
+            "format": "gemma4",
+            "config": {
+                "type": "gpt",
+                "base_model": {
+                    "decoder": {
+                        "type": "pattern",
+                        "blocks": {
+                            "full_attention": {
+                                "type": "decoder",
+                                "mixer": {
+                                    "type": "attention",
+                                    "rotary": {
+                                        "type": "proportional",
+                                        "theta": 1000000.0,
+                                        "partial_rotary_factor": 0.25
+                                    },
+                                    "heads": 16,
+                                    "head_groups": 2,
+                                    "head_size": 512,
+                                    "add_linear_biases": false,
+                                    "dropout": 0.0,
+                                    "softmax_scale_power": 0.0,
+                                    "implementation": "sdpa",
+                                    "query_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "key_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "value_norm": {
+                                        "type": "fixed_rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "shared_key_value": true
+                                },
+                                "mlp": {
+                                    "type": "hybrid_moe",
+                                    "dense": {
+                                        "type": "mlp",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 2112,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full"
+                                    },
+                                    "routed": {
+                                        "type": "moe",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 704,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full",
+                                        "router_normalization": {
+                                            "type": "fixed_rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "router_scale": {
+                                            "enabled": true
+                                        },
+                                        "router_input_scale": 0.018844459036110227,
+                                        "router_per_expert_scale": {
+                                            "enabled": true
+                                        },
+                                        "experts": 128,
+                                        "experts_per_token": 8,
+                                        "auxiliary_loss_coefficient": 0.0
+                                    }
+                                },
+                                "normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "pre_mlp_normalization": {
+                                    "type": "none"
+                                },
+                                "post_mixer_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "post_mlp_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "output_scale": {
+                                    "lr_scale": 0.0,
+                                    "enabled": true
+                                }
+                            },
+                            "sliding_attention": {
+                                "type": "decoder",
+                                "mixer": {
+                                    "type": "attention",
+                                    "rotary": {
+                                        "type": "default",
+                                        "theta": 10000.0
+                                    },
+                                    "heads": 16,
+                                    "head_groups": 8,
+                                    "head_size": 256,
+                                    "add_linear_biases": false,
+                                    "dropout": 0.0,
+                                    "window_size": 1024,
+                                    "softmax_scale_power": 0.0,
+                                    "implementation": "flash",
+                                    "query_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "key_norm": {
+                                        "type": "rms_norm",
+                                        "epsilon": 1e-06
+                                    },
+                                    "value_norm": {
+                                        "type": "fixed_rms_norm",
+                                        "epsilon": 1e-06
+                                    }
+                                },
+                                "mlp": {
+                                    "type": "hybrid_moe",
+                                    "dense": {
+                                        "type": "mlp",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 2112,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full"
+                                    },
+                                    "routed": {
+                                        "type": "moe",
+                                        "pre_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "post_norm": {
+                                            "type": "rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "intermediate_size": 704,
+                                        "add_linear_biases": false,
+                                        "gated": true,
+                                        "activation": "gelu",
+                                        "recompute_level": "full",
+                                        "router_normalization": {
+                                            "type": "fixed_rms_norm",
+                                            "epsilon": 1e-06
+                                        },
+                                        "router_scale": {
+                                            "enabled": true
+                                        },
+                                        "router_input_scale": 0.018844459036110227,
+                                        "router_per_expert_scale": {
+                                            "enabled": true
+                                        },
+                                        "experts": 128,
+                                        "experts_per_token": 8,
+                                        "auxiliary_loss_coefficient": 0.0
+                                    }
+                                },
+                                "normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "pre_mlp_normalization": {
+                                    "type": "none"
+                                },
+                                "post_mixer_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "post_mlp_normalization": {
+                                    "type": "rms_norm",
+                                    "epsilon": 1e-06
+                                },
+                                "output_scale": {
+                                    "lr_scale": 0.0,
+                                    "enabled": true
+                                }
+                            }
+                        },
+                        "pattern": [
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "sliding_attention",
+                            "full_attention"
+                        ],
+                        "num_blocks": 30
+                    },
+                    "embeddings": {
+                        "vocab_size": 262144,
+                        "vocab_parallel": false,
+                        "embedding_scale": 53.0659966456864
+                    },
+                    "head": {
+                        "normalization": {
+                            "type": "rms_norm",
+                            "epsilon": 1e-06
+                        },
+                        "cross_entropy_splits": 8,
+                        "final_logit_softcap": 30.0
+                    },
+                    "tied_embedding_weight": true,
+                    "hidden_size": 2816,
+                    "peft": {
+                        "type": "none"
+                    }
+                },
+                "multi_stage": {
+                    "zero_stage": 3
+                },
+                "distributed": {
+                    "tensor_parallel": 2,
+                    "sequence_tensor_parallel": true,
+                    "sequence_data_parallel": 4,
+                    "world_size": 32,
+                    "rank": 0,
+                    "local_world_size": 8,
+                    "timeout": 3600.0,
+                    "compute_dtype": "bfloat16"
+                }
+            },
+            "shards": [
+                "weights"
+            ],
+            "metadata": {
+                "optimizer": {
+                    "current_step": 70,
+                    "grad_scaler": {
+                        "type": "NoopGradScaler"
+                    }
+                },
+                "completed_steps": 70,
+                "metrics": {
+                    "training": {
+                        "batch_size": 2097152,
+                        "load_balancing_loss": 823.45361328125,
+                        "lm_head_loss": 0.15507733821868896,
+                        "total_steps": 210,
+                        "completed_steps": 70,
+                        "consumed_tokens": 146800640,
+                        "percent_done": 33.333333333333336,
+                        "step_time_ms": 27782.785279676318,
+                        "step_time_average_ms": 28139.88637443898,
+                        "remaining_time": 3939.584092421457,
+                        "completion_time": 1788872635.9465992,
+                        "skipped_iters": 0,
+                        "nan_iters": 0,
+                        "model_tflops": 60428.45211557096,
+                        "hardware_tflops": 74970.05489154597,
+                        "tokens_per_sec_per_gpu": 2358.870766205753,
+                        "run": 0,
+                        "grad_norm": 4.336630344390869,
+                        "learning_rate": 1.5611870653623826e-05,
+                        "loss_scale": 1.0,
+                        "reserved": 61408.0,
+                        "allocated": 21478.49755859375,
+                        "max_allocated": 54787.2607421875,
+                        "max_reserved": 61408.0,
+                        "global_max_reserved": 61408.0
+                    }
+                }
+            }
+        },
+        "model_config": {
+            "vocab_size": 262144,
+            "num_attention_heads": 16,
+            "num_key_value_heads": 8,
+            "head_dim": 256,
+            "global_head_dim": 512,
+            "num_global_key_value_heads": 2,
+            "attention_bias": false,
+            "attention_dropout": 0.0,
+            "sliding_window": 1024,
+            "rms_norm_eps": 1e-06,
+            "attention_k_eq_v": true,
+            "rope_parameters": {
+                "sliding_attention": {
+                    "rope_type": "default",
+                    "rope_theta": 10000.0
+                },
+                "full_attention": {
+                    "rope_type": "proportional",
+                    "rope_theta": 1000000.0,
+                    "partial_rotary_factor": 0.25
+                }
+            },
+            "intermediate_size": 2112,
+            "hidden_activation": "gelu_pytorch_tanh",
+            "num_experts": 128,
+            "top_k_experts": 8,
+            "moe_intermediate_size": 704,
+            "enable_moe_block": true,
+            "num_hidden_layers": 30,
+            "layer_types": [
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "sliding_attention",
+                "full_attention"
+            ],
+            "final_logit_softcapping": 30.0,
+            "tie_word_embeddings": true,
+            "hidden_size": 2816,
+            "hidden_size_per_layer_input": 0,
+            "use_bidirectional_attention": null,
+            "model_type": "gemma4_text",
+            "architectures": [
+                "Gemma4ForCausalLM"
+            ]
+        },
+        "format": "pt"
+    },
+    "weight_map": {
+        "model.embed_tokens.weight": "model_0.safetensors",
+        "model.layers.0.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.0.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.0.router.scale": "model_0.safetensors",
+        "model.layers.0.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.0.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.0.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.0.router.proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.0.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.0.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.0.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.0.experts.down_proj": "model_0.safetensors",
+        "model.layers.0.layer_scalar": "model_0.safetensors",
+        "model.layers.1.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.1.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.1.router.scale": "model_0.safetensors",
+        "model.layers.1.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.1.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.1.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.1.router.proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.1.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.1.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.1.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.1.experts.down_proj": "model_0.safetensors",
+        "model.layers.1.layer_scalar": "model_0.safetensors",
+        "model.layers.2.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.2.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.2.router.scale": "model_0.safetensors",
+        "model.layers.2.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.2.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.2.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.2.router.proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.2.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.2.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.2.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.2.experts.down_proj": "model_0.safetensors",
+        "model.layers.2.layer_scalar": "model_0.safetensors",
+        "model.layers.3.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.3.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.3.router.scale": "model_0.safetensors",
+        "model.layers.3.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.3.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.3.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.3.router.proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.3.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.3.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.3.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.3.experts.down_proj": "model_0.safetensors",
+        "model.layers.3.layer_scalar": "model_0.safetensors",
+        "model.layers.4.input_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_attention_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.q_norm.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.k_norm.weight": "model_0.safetensors",
+        "model.layers.4.pre_feedforward_layernorm.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm_1.weight": "model_0.safetensors",
+        "model.layers.4.router.scale": "model_0.safetensors",
+        "model.layers.4.router.per_expert_scale": "model_0.safetensors",
+        "model.layers.4.pre_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.4.post_feedforward_layernorm_2.weight": "model_0.safetensors",
+        "model.layers.4.router.proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.q_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.k_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.v_proj.weight": "model_0.safetensors",
+        "model.layers.4.self_attn.o_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.gate_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.up_proj.weight": "model_0.safetensors",
+        "model.layers.4.mlp.down_proj.weight": "model_0.safetensors",
+        "model.layers.4.experts.gate_up_proj": "model_0.safetensors",
+        "model.layers.4.experts.down_proj": "model_1.safetensors",
+        "model.layers.4.layer_scalar": "model_1.safetensors",
+        "model.layers.5.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.5.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.5.router.scale": "model_1.safetensors",
+        "model.layers.5.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.5.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.5.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.5.router.proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.5.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.5.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.5.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.5.experts.down_proj": "model_1.safetensors",
+        "model.layers.5.layer_scalar": "model_1.safetensors",
+        "model.layers.6.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.6.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.6.router.scale": "model_1.safetensors",
+        "model.layers.6.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.6.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.6.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.6.router.proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.6.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.6.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.6.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.6.experts.down_proj": "model_1.safetensors",
+        "model.layers.6.layer_scalar": "model_1.safetensors",
+        "model.layers.7.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.7.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.7.router.scale": "model_1.safetensors",
+        "model.layers.7.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.7.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.7.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.7.router.proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.7.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.7.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.7.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.7.experts.down_proj": "model_1.safetensors",
+        "model.layers.7.layer_scalar": "model_1.safetensors",
+        "model.layers.8.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.8.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.8.router.scale": "model_1.safetensors",
+        "model.layers.8.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.8.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.8.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.8.router.proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.8.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.8.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.8.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.8.experts.down_proj": "model_1.safetensors",
+        "model.layers.8.layer_scalar": "model_1.safetensors",
+        "model.layers.9.input_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_attention_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.q_norm.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.k_norm.weight": "model_1.safetensors",
+        "model.layers.9.pre_feedforward_layernorm.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm_1.weight": "model_1.safetensors",
+        "model.layers.9.router.scale": "model_1.safetensors",
+        "model.layers.9.router.per_expert_scale": "model_1.safetensors",
+        "model.layers.9.pre_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.9.post_feedforward_layernorm_2.weight": "model_1.safetensors",
+        "model.layers.9.router.proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.q_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.k_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.v_proj.weight": "model_1.safetensors",
+        "model.layers.9.self_attn.o_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.gate_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.up_proj.weight": "model_1.safetensors",
+        "model.layers.9.mlp.down_proj.weight": "model_1.safetensors",
+        "model.layers.9.experts.gate_up_proj": "model_1.safetensors",
+        "model.layers.9.experts.down_proj": "model_1.safetensors",
+        "model.layers.9.layer_scalar": "model_2.safetensors",
+        "model.layers.10.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.10.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.10.router.scale": "model_2.safetensors",
+        "model.layers.10.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.10.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.10.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.10.router.proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.10.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.10.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.10.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.10.experts.down_proj": "model_2.safetensors",
+        "model.layers.10.layer_scalar": "model_2.safetensors",
+        "model.layers.11.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.11.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.11.router.scale": "model_2.safetensors",
+        "model.layers.11.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.11.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.11.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.11.router.proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.11.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.11.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.11.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.11.experts.down_proj": "model_2.safetensors",
+        "model.layers.11.layer_scalar": "model_2.safetensors",
+        "model.layers.12.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.12.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.12.router.scale": "model_2.safetensors",
+        "model.layers.12.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.12.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.12.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.12.router.proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.12.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.12.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.12.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.12.experts.down_proj": "model_2.safetensors",
+        "model.layers.12.layer_scalar": "model_2.safetensors",
+        "model.layers.13.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.13.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.13.router.scale": "model_2.safetensors",
+        "model.layers.13.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.13.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.13.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.13.router.proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.13.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.13.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.13.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.13.experts.down_proj": "model_2.safetensors",
+        "model.layers.13.layer_scalar": "model_2.safetensors",
+        "model.layers.14.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.14.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.14.router.scale": "model_2.safetensors",
+        "model.layers.14.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.14.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.14.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.14.router.proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.14.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.14.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.14.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.14.experts.down_proj": "model_2.safetensors",
+        "model.layers.14.layer_scalar": "model_2.safetensors",
+        "model.layers.15.input_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_attention_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.q_norm.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.k_norm.weight": "model_2.safetensors",
+        "model.layers.15.pre_feedforward_layernorm.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm_1.weight": "model_2.safetensors",
+        "model.layers.15.router.scale": "model_2.safetensors",
+        "model.layers.15.router.per_expert_scale": "model_2.safetensors",
+        "model.layers.15.pre_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.15.post_feedforward_layernorm_2.weight": "model_2.safetensors",
+        "model.layers.15.router.proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.q_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.k_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.v_proj.weight": "model_2.safetensors",
+        "model.layers.15.self_attn.o_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.gate_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.up_proj.weight": "model_2.safetensors",
+        "model.layers.15.mlp.down_proj.weight": "model_2.safetensors",
+        "model.layers.15.experts.gate_up_proj": "model_2.safetensors",
+        "model.layers.15.experts.down_proj": "model_3.safetensors",
+        "model.layers.15.layer_scalar": "model_3.safetensors",
+        "model.layers.16.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.16.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.16.router.scale": "model_3.safetensors",
+        "model.layers.16.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.16.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.16.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.16.router.proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.16.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.16.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.16.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.16.experts.down_proj": "model_3.safetensors",
+        "model.layers.16.layer_scalar": "model_3.safetensors",
+        "model.layers.17.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.17.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.17.router.scale": "model_3.safetensors",
+        "model.layers.17.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.17.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.17.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.17.router.proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.17.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.17.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.17.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.17.experts.down_proj": "model_3.safetensors",
+        "model.layers.17.layer_scalar": "model_3.safetensors",
+        "model.layers.18.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.18.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.18.router.scale": "model_3.safetensors",
+        "model.layers.18.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.18.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.18.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.18.router.proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.18.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.18.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.18.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.18.experts.down_proj": "model_3.safetensors",
+        "model.layers.18.layer_scalar": "model_3.safetensors",
+        "model.layers.19.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.19.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.19.router.scale": "model_3.safetensors",
+        "model.layers.19.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.19.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.19.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.19.router.proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.19.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.19.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.19.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.19.experts.down_proj": "model_3.safetensors",
+        "model.layers.19.layer_scalar": "model_3.safetensors",
+        "model.layers.20.input_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_attention_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.q_norm.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.k_norm.weight": "model_3.safetensors",
+        "model.layers.20.pre_feedforward_layernorm.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm_1.weight": "model_3.safetensors",
+        "model.layers.20.router.scale": "model_3.safetensors",
+        "model.layers.20.router.per_expert_scale": "model_3.safetensors",
+        "model.layers.20.pre_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.20.post_feedforward_layernorm_2.weight": "model_3.safetensors",
+        "model.layers.20.router.proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.q_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.k_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.v_proj.weight": "model_3.safetensors",
+        "model.layers.20.self_attn.o_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.gate_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.up_proj.weight": "model_3.safetensors",
+        "model.layers.20.mlp.down_proj.weight": "model_3.safetensors",
+        "model.layers.20.experts.gate_up_proj": "model_3.safetensors",
+        "model.layers.20.experts.down_proj": "model_3.safetensors",
+        "model.layers.20.layer_scalar": "model_4.safetensors",
+        "model.layers.21.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.21.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.21.router.scale": "model_4.safetensors",
+        "model.layers.21.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.21.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.21.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.21.router.proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.21.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.21.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.21.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.21.experts.down_proj": "model_4.safetensors",
+        "model.layers.21.layer_scalar": "model_4.safetensors",
+        "model.layers.22.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.22.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.22.router.scale": "model_4.safetensors",
+        "model.layers.22.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.22.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.22.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.22.router.proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.22.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.22.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.22.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.22.experts.down_proj": "model_4.safetensors",
+        "model.layers.22.layer_scalar": "model_4.safetensors",
+        "model.layers.23.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.23.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.23.router.scale": "model_4.safetensors",
+        "model.layers.23.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.23.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.23.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.23.router.proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.23.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.23.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.23.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.23.experts.down_proj": "model_4.safetensors",
+        "model.layers.23.layer_scalar": "model_4.safetensors",
+        "model.layers.24.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.24.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.24.router.scale": "model_4.safetensors",
+        "model.layers.24.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.24.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.24.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.24.router.proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.24.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.24.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.24.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.24.experts.down_proj": "model_4.safetensors",
+        "model.layers.24.layer_scalar": "model_4.safetensors",
+        "model.layers.25.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.25.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.25.router.scale": "model_4.safetensors",
+        "model.layers.25.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.25.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.25.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.25.router.proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.25.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.25.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.25.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.25.experts.down_proj": "model_4.safetensors",
+        "model.layers.25.layer_scalar": "model_4.safetensors",
+        "model.layers.26.input_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_attention_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.q_norm.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.k_norm.weight": "model_4.safetensors",
+        "model.layers.26.pre_feedforward_layernorm.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm_1.weight": "model_4.safetensors",
+        "model.layers.26.router.scale": "model_4.safetensors",
+        "model.layers.26.router.per_expert_scale": "model_4.safetensors",
+        "model.layers.26.pre_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.26.post_feedforward_layernorm_2.weight": "model_4.safetensors",
+        "model.layers.26.router.proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.q_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.k_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.v_proj.weight": "model_4.safetensors",
+        "model.layers.26.self_attn.o_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.gate_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.up_proj.weight": "model_4.safetensors",
+        "model.layers.26.mlp.down_proj.weight": "model_4.safetensors",
+        "model.layers.26.experts.gate_up_proj": "model_4.safetensors",
+        "model.layers.26.experts.down_proj": "model_5.safetensors",
+        "model.layers.26.layer_scalar": "model_5.safetensors",
+        "model.layers.27.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.27.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.27.router.scale": "model_5.safetensors",
+        "model.layers.27.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.27.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.27.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.27.router.proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.v_proj.weight": "model_5.safetensors",
+        "model.layers.27.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.27.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.27.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.27.experts.down_proj": "model_5.safetensors",
+        "model.layers.27.layer_scalar": "model_5.safetensors",
+        "model.layers.28.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.28.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.28.router.scale": "model_5.safetensors",
+        "model.layers.28.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.28.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.28.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.28.router.proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.v_proj.weight": "model_5.safetensors",
+        "model.layers.28.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.28.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.28.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.28.experts.down_proj": "model_5.safetensors",
+        "model.layers.28.layer_scalar": "model_5.safetensors",
+        "model.layers.29.input_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_attention_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.q_norm.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.k_norm.weight": "model_5.safetensors",
+        "model.layers.29.pre_feedforward_layernorm.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm_1.weight": "model_5.safetensors",
+        "model.layers.29.router.scale": "model_5.safetensors",
+        "model.layers.29.router.per_expert_scale": "model_5.safetensors",
+        "model.layers.29.pre_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.29.post_feedforward_layernorm_2.weight": "model_5.safetensors",
+        "model.layers.29.router.proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.q_proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.k_proj.weight": "model_5.safetensors",
+        "model.layers.29.self_attn.o_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.gate_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.up_proj.weight": "model_5.safetensors",
+        "model.layers.29.mlp.down_proj.weight": "model_5.safetensors",
+        "model.layers.29.experts.gate_up_proj": "model_5.safetensors",
+        "model.layers.29.experts.down_proj": "model_5.safetensors",
+        "model.layers.29.layer_scalar": "model_5.safetensors",
+        "model.norm.weight": "model_5.safetensors"
+    }
+}

26_09_08_sft_email_v4_v3/export/gemma4/70/model_0.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d193154bc227e9b768dae5398a411cc8fe1d490f8eba3df763ea15805c9f517
+size 18219689440

26_09_08_sft_email_v4_v3/export/gemma4/70/model_1.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e036e96a3b276aea68fea94af6dfe1b109dca14d20c43f877a25dac09cc00f5b
+size 17354616388

26_09_08_sft_email_v4_v3/export/gemma4/70/model_2.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b5ffba35f5720f0136ad3515f9eaac87883b0b498bbb037a0b01cab0e3816bd
+size 18580955576

26_09_08_sft_email_v4_v3/export/gemma4/70/model_3.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81bbb0af18418e2e9606acc0f6366e2b8e58b5a2068b0315b872f22fd04d6920
+size 17354616500

26_09_08_sft_email_v4_v3/export/gemma4/70/model_4.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6654be93b2273f19ce97c74ae00bddbbf25abb1d178edb8cd5354d3ebb277a69
+size 18580955552

26_09_08_sft_email_v4_v3/export/gemma4/70/model_5.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a792bbb9303b8519c827e824db7aaa34e8932ee67b1b18c17832f74ec492942c
+size 10841863000