| { |
| "analysis_sha256": "21c80c98be1576ed3f52cd6be881bb9cd7e67ef50a3e830e23d155ceaebfacee", |
| "architecture_profile": { |
| "adapter_id": "mistral-devstral-dense-v1", |
| "config_model_type": "mistral", |
| "dense": true, |
| "mtp_declared": false, |
| "notes": [ |
| "Mistral/Devstral dense text checkpoints (model_type=mistral|llama).", |
| "MLX-LM remaps model_type=mistral to the llama implementation." |
| ], |
| "optimization_scope": "text-path", |
| "product_family": "mistral-devstral", |
| "support_level": "supported", |
| "support_tier": "convertible", |
| "text_layer_count": 40, |
| "vision_present": false |
| }, |
| "assignments": [ |
| { |
| "bits": 8, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.0253125, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.045000000000000005, |
| "long_context_loss": 0.016875, |
| "mtp_acceptance_loss": 0.0045000000000000005, |
| "output_kl": 0.05625, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.036562500000000005, |
| "token_disagreement": 0.030937500000000003 |
| }, |
| "module_path": "model.embed_tokens", |
| "outlier_strategy": "none", |
| "parameters": 671088640, |
| "predicted_loss": 0.0307125, |
| "reason": "protected embedding policy", |
| "role": "embedding", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 3, |
| "storage_bpw": 8.5 |
| }, |
| "tensor": "model.embed_tokens.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.0.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.0.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.0.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.0.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.0.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.0.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.0.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.0.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.0.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.0.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.0.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.0.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.0.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.0.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.0.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.0.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.0.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.0.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.1.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.1.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.1.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.1.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.1.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.1.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.1.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.1.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.1.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.1.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.1.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.1.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.1.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.1.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.1.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.1.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.1.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.1.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.2.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.2.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.2.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.2.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.2.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.2.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.2.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.2.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.2.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.2.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.2.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.2.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.2.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.2.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.2.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.2.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.2.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.2.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.3.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.3.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.3.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.3.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.3.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.3.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.3.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.3.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.3.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.3.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.3.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.3.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.3.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.3.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.3.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.3.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.3.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.3.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.4.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.4.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.4.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.4.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.4.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.4.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.4.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.4.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.4.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.4.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.4.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.4.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.4.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.4.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.4.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.4.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.4.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.4.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.5.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.5.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.5.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.5.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.5.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.5.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.5.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.5.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.5.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.5.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.5.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.5.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.5.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.5.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.5.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.5.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.5.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.5.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.6.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.6.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.6.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.6.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.6.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.6.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.6.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.6.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.6.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.6.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.6.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.6.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.6.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.6.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.6.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.6.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.6.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.6.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.7.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.7.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.7.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.7.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.7.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.7.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.7.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.7.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.7.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.7.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.10.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.10.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.10.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.10.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.10.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.10.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.10.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.10.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.10.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.10.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.10.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.10.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.10.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.10.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.10.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.10.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.10.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.10.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.11.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.11.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.11.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.11.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.11.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.11.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.11.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.11.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.11.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.11.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.11.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.11.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.7.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.7.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.7.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.7.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.7.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.7.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.7.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.7.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.8.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.8.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.8.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.8.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.8.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.8.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.8.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.8.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.8.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.8.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.8.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.8.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.8.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.8.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.8.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.8.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.8.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.8.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.9.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.9.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.9.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.9.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.9.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.9.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.9.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.9.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.9.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.9.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.9.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.9.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.9.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.9.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.9.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.9.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.9.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.9.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.11.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.11.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.11.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.11.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.11.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.11.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.12.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.12.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.12.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.12.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.12.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.12.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.12.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.12.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.12.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.12.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.12.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.12.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.12.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.12.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.12.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.12.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.12.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.12.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.13.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.13.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.13.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.13.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.13.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.13.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.13.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.13.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.13.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.13.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.13.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.13.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.13.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.13.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.13.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.13.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.13.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.13.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.14.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.14.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.14.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.14.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.14.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.14.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.14.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.14.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.14.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.14.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.14.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.14.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.14.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.14.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.14.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.14.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.14.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.14.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.15.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.15.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.15.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.15.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.15.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.15.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.15.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.15.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.15.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.15.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.15.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.15.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.15.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.15.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.15.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.15.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.15.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.15.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.16.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.16.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.16.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.16.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.16.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.16.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.16.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.16.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.16.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.16.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.16.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.16.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.16.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.16.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.16.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.16.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.16.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.16.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.17.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.17.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.17.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.17.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.17.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.17.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.17.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.17.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.17.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.17.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.17.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.17.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.17.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.17.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.17.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.17.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.17.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.17.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.18.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.18.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.18.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.18.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.18.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.18.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.18.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.18.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.18.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.18.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.18.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.18.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.18.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.18.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.18.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.18.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.18.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.18.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.19.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.19.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.19.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.19.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.19.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.19.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.19.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.19.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.19.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.19.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.19.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.19.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.19.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.19.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.19.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.19.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.19.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.19.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.20.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.20.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.20.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.20.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.20.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.20.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.20.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.20.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.20.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.20.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.20.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.20.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.20.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.20.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.20.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.20.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.20.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.20.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.21.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.21.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.21.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.21.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.21.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.21.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.21.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.21.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.21.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.21.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.21.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.21.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.21.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.21.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.21.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.21.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.21.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.21.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.22.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.22.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.22.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.22.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.22.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.22.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.22.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.22.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.22.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.22.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.22.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.22.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.22.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.22.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.22.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.22.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.22.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.22.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.23.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.23.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.23.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.23.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.23.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.23.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.23.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.23.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.23.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.23.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.23.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.23.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.23.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.23.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.23.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.23.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.23.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.23.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.24.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.24.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.24.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.24.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.24.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.24.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.24.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.24.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.24.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.24.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.24.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.24.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.24.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.24.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.24.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.24.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.24.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.24.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.25.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.25.input_layernorm.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.25.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.25.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.25.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.25.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 4, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.14318912319027588, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.25455844122715715, |
| "long_context_loss": 0.09545941546018392, |
| "mtp_acceptance_loss": 0.025455844122715714, |
| "output_kl": 0.3181980515339464, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.20682873349706518, |
| "token_disagreement": 0.17500892834367054 |
| }, |
| "module_path": "model.layers.25.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.17373613613753475, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 5.0 |
| }, |
| "tensor": "model.layers.25.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.25.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.25.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.25.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.25.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.25.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.25.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.25.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.25.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.25.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.25.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.26.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.26.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.26.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.26.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.26.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.26.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.26.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.26.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.26.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.26.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.26.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.26.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.27.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.27.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.27.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.27.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.27.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.27.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.27.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.27.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.27.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.27.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.27.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.27.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.28.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.28.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.28.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.28.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.28.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.28.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.28.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.28.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.28.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.28.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.28.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.28.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.29.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.29.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.29.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.29.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.29.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.29.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.29.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.29.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.29.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.29.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.29.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.29.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.30.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.30.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.30.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.30.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.30.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.30.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.30.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.30.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.30.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.30.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.30.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.30.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.31.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.31.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.31.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.31.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.31.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.31.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.31.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.31.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.31.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.31.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.31.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.31.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.32.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.32.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.32.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.32.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.32.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.32.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.32.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.32.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.32.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.32.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.32.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.32.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.33.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.33.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.33.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.33.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.33.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.33.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.33.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.33.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.33.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.33.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.33.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.33.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.34.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.34.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.34.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.34.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.34.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.34.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.34.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.34.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.34.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.34.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.34.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.34.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.35.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.35.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.35.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.35.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.35.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.35.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.35.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.35.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.35.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.35.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.35.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.35.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.081, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.144, |
| "long_context_loss": 0.22499999999999998, |
| "mtp_acceptance_loss": 0.0144, |
| "output_kl": 0.18, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.11699999999999999, |
| "token_disagreement": 0.099 |
| }, |
| "module_path": "model.layers.35.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.11195999999999999, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.35.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.35.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.35.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.35.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.35.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.36.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.36.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.36.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.36.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.36.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.36.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.36.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.36.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.36.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.36.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.36.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.36.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.36.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.36.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.36.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.36.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.36.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.36.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.37.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.37.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.37.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.37.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.37.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.37.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.37.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.37.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.37.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.37.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.37.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.37.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "lm_head", |
| "outlier_strategy": "none", |
| "parameters": 671088640, |
| "predicted_loss": 0.0, |
| "reason": "protected lm_head policy", |
| "role": "lm_head", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "lm_head.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.37.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.37.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.37.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.37.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.37.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.37.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.38.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.38.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.38.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.38.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.38.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.38.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.38.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.38.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.38.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.38.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.38.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.38.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.38.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.38.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.38.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.38.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.38.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.38.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.39.input_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.39.input_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.39.mlp.down_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.39.mlp.down_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.39.mlp.gate_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.39.mlp.gate_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 64, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.050625, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.09000000000000001, |
| "long_context_loss": 0.03375, |
| "mtp_acceptance_loss": 0.009000000000000001, |
| "output_kl": 0.1125, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.07312500000000001, |
| "token_disagreement": 0.061875000000000006 |
| }, |
| "module_path": "model.layers.39.mlp.up_proj", |
| "outlier_strategy": "none", |
| "parameters": 167772160, |
| "predicted_loss": 0.061425, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "mlp", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 6.5 |
| }, |
| "tensor": "model.layers.39.mlp.up_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.layers.39.post_attention_layernorm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.layers.39.post_attention_layernorm.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.39.self_attn.k_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.39.self_attn.k_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.39.self_attn.o_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.39.self_attn.o_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.39.self_attn.q_proj", |
| "outlier_strategy": "none", |
| "parameters": 20971520, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.39.self_attn.q_proj.weight" |
| }, |
| { |
| "bits": 6, |
| "group_size": 32, |
| "method": "affine", |
| "metrics": { |
| "cosine_distance": 0.057275649276110355, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.10182337649086284, |
| "long_context_loss": 0.15909902576697318, |
| "mtp_acceptance_loss": 0.010182337649086284, |
| "output_kl": 0.12727922061357855, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.08273149339882606, |
| "token_disagreement": 0.0700035713374682 |
| }, |
| "module_path": "model.layers.39.self_attn.v_proj", |
| "outlier_strategy": "none", |
| "parameters": 5242880, |
| "predicted_loss": 0.07916767522164586, |
| "reason": "selected by marginal quality gain per storage bit", |
| "role": "attention", |
| "scale_strategy": "group-affine", |
| "strategy_metadata": { |
| "selected_from_candidates": 7, |
| "storage_bpw": 7.0 |
| }, |
| "tensor": "model.layers.39.self_attn.v_proj.weight" |
| }, |
| { |
| "bits": 16, |
| "group_size": null, |
| "method": "bf16", |
| "metrics": { |
| "cosine_distance": 0.0, |
| "decode_latency_cost": 0.0, |
| "hidden_state_error": 0.0, |
| "long_context_loss": 0.0, |
| "mtp_acceptance_loss": 0.0, |
| "output_kl": 0.0, |
| "peak_memory_cost": 0.0, |
| "prefill_latency_cost": 0.0, |
| "task_loss_delta": 0.0, |
| "token_disagreement": 0.0 |
| }, |
| "module_path": "model.norm", |
| "outlier_strategy": "none", |
| "parameters": 5120, |
| "predicted_loss": 0.0, |
| "reason": "non-quantizable tensor preserved", |
| "role": "norm", |
| "scale_strategy": "none", |
| "strategy_metadata": { |
| "selected_from_candidates": 1, |
| "storage_bpw": 16.0 |
| }, |
| "tensor": "model.norm.weight" |
| } |
| ], |
| "calibration": null, |
| "candidate_bits": [ |
| 4, |
| 6, |
| 8, |
| 16 |
| ], |
| "candidate_group_sizes": [ |
| 32, |
| 64 |
| ], |
| "constraints": { |
| "effective_bpw_limit": 6.0, |
| "lm_head_min_bits": 16, |
| "max_model_size_ratio_to_uniform4": 1.1, |
| "minimum_mtp_acceptance_retention": 0.95, |
| "minimum_mtp_speedup": 1.2, |
| "minimum_quality_retention": 0.98 |
| }, |
| "created_at": "2026-08-02T18:31:06.189264Z", |
| "effective_bpw": 5.999953518527971, |
| "evidence_kind": "architecture_prior", |
| "global_validation_required": true, |
| "group_size": 32, |
| "hardware": { |
| "name": "ax-engine-apple-silicon-affine-dwq-v2", |
| "runtime": "ax-engine", |
| "supported_bits": [ |
| 2, |
| 3, |
| 4, |
| 6, |
| 8, |
| 16 |
| ], |
| "supported_group_sizes": [ |
| 32, |
| 64, |
| 128 |
| ], |
| "supported_methods": [ |
| "affine", |
| "awq", |
| "dwq", |
| "bf16" |
| ] |
| }, |
| "kv_cache": null, |
| "mtp": { |
| "candidate_bits": [ |
| 8, |
| 16 |
| ], |
| "min_bits": 8, |
| "mode": "protected", |
| "optimize_for_acceptance": true, |
| "preserve_external_sidecar": true, |
| "protect_norms": true, |
| "protect_output_head": true |
| }, |
| "mtp_distribution": {}, |
| "nominal_bpw": 5.231506184316413, |
| "objective": { |
| "cosine_distance": 0.05, |
| "decode_latency_cost": 0.05, |
| "hidden_state_error": 0.12, |
| "long_context_loss": 0.08, |
| "mtp_acceptance_loss": 0.15, |
| "output_kl": 0.25, |
| "peak_memory_cost": 0.04, |
| "prefill_latency_cost": 0.03, |
| "task_loss_delta": 0.15, |
| "token_disagreement": 0.08 |
| }, |
| "primary_runtime": "ax-engine", |
| "profile": "general", |
| "quantizer": "axquant", |
| "random_seed": 0, |
| "schema_version": "axquant.plan.v1", |
| "software_versions": { |
| "ax_engine": null, |
| "axquant": "1.0.0", |
| "mlx": "0.32.0", |
| "mlx_lm": "0.31.3", |
| "pydantic": "2.13.4", |
| "python": "3.13.14", |
| "safetensors": "0.8.0" |
| }, |
| "source_model": { |
| "architecture": "MistralForCausalLM", |
| "format": "mlx", |
| "local_path": null, |
| "model_id": "mistralai/Devstral-Small-2505", |
| "revision": null |
| }, |
| "status": "planned", |
| "target_bpw": 6.0, |
| "target_class": "4bit", |
| "target_mode": "low-memory", |
| "warnings": [ |
| "This report contains architecture priors, not calibration measurements.", |
| "Conversion planning requires --allow-unmeasured for this report.", |
| "Multi-group architecture priors weakly prefer smaller group sizes; this is development evidence only (AXQ-028).", |
| "Plan uses non-release architecture_prior evidence and requires complete-model validation.", |
| "convert ladder: prior" |
| ], |
| "weight_distribution": { |
| "4bit": { |
| "fraction": 0.5551503751641241, |
| "parameters": 13086228480 |
| }, |
| "6bit": { |
| "fraction": 0.38789353136467647, |
| "parameters": 9143582720 |
| }, |
| "8bit": { |
| "fraction": 0.028469250008416623, |
| "parameters": 671088640 |
| }, |
| "bf16": { |
| "fraction": 0.028486843462782786, |
| "parameters": 671503360 |
| } |
| } |
| } |
|
|
