{ "analysis_sha256": "e0b2d1e3ea2e95a9b7e811fa118896da76e7306af1240819dfb84dbf1cf8f44a", "architecture_profile": { "adapter_id": "qwen35-moe-v1", "audio_present": false, "config_model_type": "qwen3_5_moe", "dense": false, "mtp_declared": false, "notes": [ "AXQuant optimizes the Qwen3.5-class 35B-A3B MoE language path only (development convert; not the Qwen 3.6 certification track).", "Vision tensors are preserved at BF16 and VLM quality is not claimed.", "Fine-tunes such as Ornith-1.0-35B are eligible when the text_config matches the 35B-A3B MoE signature.", "MoE experts quantize as fused switch modules with uniform per-group precision; label packs as development evidence until certified." ], "optimization_scope": "text-path", "product_family": "qwen3.5-moe", "support_level": "supported", "support_tier": "convertible", "text_layer_count": 40, "vision_present": true }, "assignments": [ { "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": 508559360, "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": 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.language_model.embed_tokens", "outlier_strategy": "none", "parameters": 508559360, "predicted_loss": 0.026555625000000003, "reason": "protected embedding policy", "role": "embedding", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_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.language_model.layers.0.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.0.input_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.language_model.layers.0.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.0.linear_attn.A_log" }, { "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.language_model.layers.0.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.0.linear_attn.conv1d.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.language_model.layers.0.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.0.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.0.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.0.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.0.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.0.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.linear_attn.in_proj_z.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.language_model.layers.0.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.0.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.0.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.0.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.0.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.0.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.0.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.0.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.0.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.0.mlp.shared_expert_gate.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.language_model.layers.0.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.0.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.language_model.layers.1.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.1.input_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.language_model.layers.1.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.1.linear_attn.A_log" }, { "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.language_model.layers.1.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.1.linear_attn.conv1d.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.language_model.layers.1.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.1.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.1.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.1.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.1.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.1.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.linear_attn.in_proj_z.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.language_model.layers.1.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.1.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.1.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.1.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.1.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.1.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.1.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.1.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.1.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.1.mlp.shared_expert_gate.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.language_model.layers.1.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.1.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.language_model.layers.2.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.2.input_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.language_model.layers.2.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.2.linear_attn.A_log" }, { "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.language_model.layers.2.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.2.linear_attn.conv1d.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.language_model.layers.2.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.2.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.2.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.2.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.2.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.2.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.linear_attn.in_proj_z.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.language_model.layers.2.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.2.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.2.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.2.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.2.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.2.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.2.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.2.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.2.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.2.mlp.shared_expert_gate.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.language_model.layers.2.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.2.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.language_model.layers.3.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.3.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.3.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.3.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.3.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.3.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.3.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.3.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.3.mlp.shared_expert_gate.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.language_model.layers.3.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.3.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.3.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.3.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.3.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.self_attn.o_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.language_model.layers.3.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.3.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.3.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.3.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_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.language_model.layers.4.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.4.input_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.language_model.layers.4.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.4.linear_attn.A_log" }, { "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.language_model.layers.4.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.4.linear_attn.conv1d.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.language_model.layers.4.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.4.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.4.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.4.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.4.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.4.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.linear_attn.in_proj_z.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.language_model.layers.4.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.4.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.4.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.4.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.4.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.4.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.4.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.4.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.4.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.4.mlp.shared_expert_gate.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.language_model.layers.4.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.4.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.language_model.layers.5.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.5.input_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.language_model.layers.5.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.5.linear_attn.A_log" }, { "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.language_model.layers.5.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.5.linear_attn.conv1d.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.language_model.layers.5.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.5.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.5.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.5.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.5.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.5.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.linear_attn.in_proj_z.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.language_model.layers.5.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.5.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.5.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.5.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.5.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.5.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.5.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.5.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.5.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.5.mlp.shared_expert_gate.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.language_model.layers.5.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.5.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.language_model.layers.6.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.6.input_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.language_model.layers.6.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.6.linear_attn.A_log" }, { "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.language_model.layers.6.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.6.linear_attn.conv1d.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.language_model.layers.6.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.6.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.6.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.6.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.6.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.6.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.linear_attn.in_proj_z.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.language_model.layers.6.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.6.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.6.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.6.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.6.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.6.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.6.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.6.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.6.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.6.mlp.shared_expert_gate.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.language_model.layers.6.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.6.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.language_model.layers.7.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.7.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.7.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.7.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.7.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.7.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.7.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.7.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.7.mlp.shared_expert_gate.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.language_model.layers.7.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.7.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.7.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.7.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.7.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.self_attn.o_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.language_model.layers.7.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.7.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.7.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.7.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_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.language_model.layers.8.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.8.input_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.language_model.layers.8.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.8.linear_attn.A_log" }, { "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.language_model.layers.8.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.8.linear_attn.conv1d.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.language_model.layers.8.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.8.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.8.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.8.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.8.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.8.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.linear_attn.in_proj_z.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.language_model.layers.8.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.8.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.8.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.8.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.8.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.8.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.8.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.8.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.8.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.8.mlp.shared_expert_gate.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.language_model.layers.8.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.8.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.language_model.layers.9.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.9.input_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.language_model.layers.9.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.9.linear_attn.A_log" }, { "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.language_model.layers.9.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.9.linear_attn.conv1d.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.language_model.layers.9.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.9.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.9.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.9.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.9.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.9.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.linear_attn.in_proj_z.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.language_model.layers.9.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.9.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.9.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.linear_attn.out_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.language_model.layers.10.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.10.input_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.language_model.layers.10.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.10.linear_attn.A_log" }, { "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.language_model.layers.10.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.10.linear_attn.conv1d.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.language_model.layers.10.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.10.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.10.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.10.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.10.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.10.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.linear_attn.in_proj_z.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.language_model.layers.10.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.10.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.10.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.10.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.10.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.10.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.10.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.10.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.10.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.10.mlp.shared_expert_gate.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.language_model.layers.10.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.10.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.language_model.layers.11.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.11.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.11.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.11.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.11.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.11.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.11.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.11.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.11.mlp.shared_expert_gate.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.language_model.layers.11.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.11.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.11.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.11.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.11.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.self_attn.o_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.language_model.layers.11.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.11.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.11.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.11.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_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.language_model.layers.12.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.12.input_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.language_model.layers.12.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.12.linear_attn.A_log" }, { "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.language_model.layers.12.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.12.linear_attn.conv1d.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.language_model.layers.12.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.12.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.12.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.12.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.12.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.12.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.linear_attn.in_proj_z.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.language_model.layers.12.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.12.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.12.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.9.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.9.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.9.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.9.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.9.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.9.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.9.mlp.shared_expert_gate.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.language_model.layers.9.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.9.post_attention_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.12.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.12.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.12.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.12.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.12.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.12.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.12.mlp.shared_expert_gate.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.language_model.layers.12.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.12.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.language_model.layers.13.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.13.input_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.language_model.layers.13.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.13.linear_attn.A_log" }, { "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.language_model.layers.13.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.13.linear_attn.conv1d.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.language_model.layers.13.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.13.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.13.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.13.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.13.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.13.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.linear_attn.in_proj_z.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.language_model.layers.13.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.13.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.13.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.13.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.13.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.13.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.13.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.13.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.13.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.13.mlp.shared_expert_gate.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.language_model.layers.13.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.13.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.language_model.layers.14.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.14.input_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.language_model.layers.14.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.14.linear_attn.A_log" }, { "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.language_model.layers.14.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.14.linear_attn.conv1d.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.language_model.layers.14.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.14.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.14.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.14.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.14.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.14.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.linear_attn.in_proj_z.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.language_model.layers.14.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.14.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.14.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.14.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.14.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.14.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.14.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.14.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.14.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.14.mlp.shared_expert_gate.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.language_model.layers.14.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.14.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.language_model.layers.15.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.15.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.15.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.15.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.15.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.15.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.15.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.15.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.15.mlp.shared_expert_gate.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.language_model.layers.15.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.15.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.language_model.layers.15.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.15.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.15.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.15.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.self_attn.o_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.language_model.layers.15.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.15.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.15.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.15.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.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.language_model.layers.16.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.16.input_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.language_model.layers.16.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.16.linear_attn.A_log" }, { "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.language_model.layers.16.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.16.linear_attn.conv1d.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.language_model.layers.16.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.16.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.16.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.16.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.16.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.16.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.linear_attn.in_proj_z.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.language_model.layers.16.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.16.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.16.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.16.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.16.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.16.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.16.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.16.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.16.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.16.mlp.shared_expert_gate.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.language_model.layers.16.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.17.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.17.input_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.language_model.layers.17.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.17.linear_attn.A_log" }, { "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.language_model.layers.17.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.17.linear_attn.conv1d.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.language_model.layers.17.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.17.linear_attn.dt_bias" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.17.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.17.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.17.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.17.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.linear_attn.in_proj_z.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.language_model.layers.17.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.17.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.17.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.17.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.17.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.17.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.17.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.17.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.17.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.17.mlp.shared_expert_gate.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.language_model.layers.17.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.17.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.language_model.layers.18.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.18.input_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.language_model.layers.18.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.18.linear_attn.A_log" }, { "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.language_model.layers.18.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.18.linear_attn.conv1d.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.language_model.layers.18.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.18.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.18.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.18.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.18.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.18.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.18.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.18.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.linear_attn.in_proj_z.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.language_model.layers.18.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.18.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.18.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.18.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.18.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.18.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.18.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.18.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.18.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.18.mlp.shared_expert_gate.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.language_model.layers.18.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.18.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.language_model.layers.19.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.19.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.19.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.19.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.19.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.19.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.19.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.19.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.19.mlp.shared_expert_gate.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.language_model.layers.19.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.19.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.language_model.layers.19.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.19.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.19.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.19.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.self_attn.o_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.language_model.layers.19.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.19.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.19.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.19.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.19.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.language_model.layers.20.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.20.input_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.language_model.layers.20.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.20.linear_attn.A_log" }, { "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.language_model.layers.20.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.20.linear_attn.conv1d.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.language_model.layers.20.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.20.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.20.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.20.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.20.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.20.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.20.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.20.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.linear_attn.in_proj_z.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.language_model.layers.20.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.20.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.20.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.20.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.20.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.20.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.20.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.20.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.20.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.20.mlp.shared_expert_gate.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.language_model.layers.20.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.21.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.21.input_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.language_model.layers.21.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.21.linear_attn.A_log" }, { "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.language_model.layers.21.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.21.linear_attn.conv1d.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.language_model.layers.21.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.21.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.21.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.21.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.21.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.21.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.21.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.21.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.linear_attn.in_proj_z.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.language_model.layers.21.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.21.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.21.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.21.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.21.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.21.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.21.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.21.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.21.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.21.mlp.shared_expert_gate.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.language_model.layers.21.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.21.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.language_model.layers.22.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.22.input_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.language_model.layers.22.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.22.linear_attn.A_log" }, { "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.language_model.layers.22.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.22.linear_attn.conv1d.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.language_model.layers.22.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.22.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.22.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.22.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.22.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.22.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.22.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.22.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.linear_attn.in_proj_z.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.language_model.layers.22.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.22.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.22.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.22.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.22.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.22.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.22.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.22.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.22.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.22.mlp.shared_expert_gate.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.language_model.layers.22.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.22.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.language_model.layers.23.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.23.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.23.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.23.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.23.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.23.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.23.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.23.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.23.mlp.shared_expert_gate.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.language_model.layers.23.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.23.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.language_model.layers.23.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.23.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.23.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.23.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.23.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.self_attn.o_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.language_model.layers.23.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.23.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.23.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.23.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.23.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.language_model.layers.24.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.24.input_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.language_model.layers.24.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.24.linear_attn.A_log" }, { "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.language_model.layers.24.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.24.linear_attn.conv1d.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.language_model.layers.24.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.24.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.24.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.24.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.24.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.24.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.24.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.24.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.linear_attn.in_proj_z.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.language_model.layers.24.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.24.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.24.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.24.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.24.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.24.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.24.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.24.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.24.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.24.mlp.shared_expert_gate.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.language_model.layers.24.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.25.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.25.input_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.language_model.layers.25.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.25.linear_attn.A_log" }, { "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.language_model.layers.25.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.25.linear_attn.conv1d.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.language_model.layers.25.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.25.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.25.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.25.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.25.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.25.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.25.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.25.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.linear_attn.in_proj_z.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.language_model.layers.25.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.25.linear_attn.norm.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.324, "decode_latency_cost": 0.0, "hidden_state_error": 0.576, "long_context_loss": 0.8999999999999999, "mtp_acceptance_loss": 0.0576, "output_kl": 0.72, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.46799999999999997, "token_disagreement": 0.396 }, "module_path": "model.language_model.layers.25.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.374112, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.25.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.25.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.25.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.25.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.25.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.25.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.25.mlp.shared_expert_gate.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.language_model.layers.25.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.25.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.language_model.layers.26.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.26.input_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.language_model.layers.26.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.26.linear_attn.A_log" }, { "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.language_model.layers.26.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.26.linear_attn.conv1d.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.language_model.layers.26.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.26.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.26.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.26.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.26.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.26.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.26.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.26.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.26.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.26.linear_attn.in_proj_z.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.language_model.layers.26.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.26.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.26.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.26.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.26.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.26.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.26.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.26.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.26.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.26.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.26.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.26.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.26.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.26.mlp.shared_expert_gate.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.language_model.layers.26.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.26.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.language_model.layers.27.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.27.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.27.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.27.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.27.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.27.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.27.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.27.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.27.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.27.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.27.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.27.mlp.shared_expert_gate.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.language_model.layers.27.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.27.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.language_model.layers.27.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.27.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.27.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.27.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.27.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.27.self_attn.o_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.language_model.layers.27.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.27.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.27.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.27.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.27.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_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.language_model.layers.28.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.28.input_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.language_model.layers.28.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.28.linear_attn.A_log" }, { "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.language_model.layers.28.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.28.linear_attn.conv1d.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.language_model.layers.28.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.28.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.28.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.28.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.28.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.28.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.28.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.28.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.28.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.28.linear_attn.in_proj_z.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.language_model.layers.28.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.28.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.28.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.28.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.28.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.28.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.28.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.28.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.28.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.28.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.28.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.28.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.28.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.28.mlp.shared_expert_gate.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.language_model.layers.28.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.28.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.language_model.layers.29.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.29.input_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.language_model.layers.29.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.29.linear_attn.A_log" }, { "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.language_model.layers.29.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.29.linear_attn.conv1d.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.language_model.layers.29.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.29.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.29.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.29.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.29.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.29.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.29.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.29.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.29.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.29.linear_attn.in_proj_z.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.language_model.layers.29.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.29.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.29.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.29.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.29.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.29.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.29.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.29.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.29.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.29.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.29.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.29.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.29.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.29.mlp.shared_expert_gate.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.language_model.layers.29.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.30.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.30.input_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.language_model.layers.30.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.30.linear_attn.A_log" }, { "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.language_model.layers.30.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.30.linear_attn.conv1d.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.language_model.layers.30.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.30.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.30.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.30.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.30.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.30.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.30.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.30.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.30.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.30.linear_attn.in_proj_z.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.language_model.layers.30.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.30.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.30.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.30.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.30.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.30.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.30.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.30.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.30.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.30.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.30.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.30.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.30.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.30.mlp.shared_expert_gate.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.language_model.layers.30.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.30.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.language_model.layers.31.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.31.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.31.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.31.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.31.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.31.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.31.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.31.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.31.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.31.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.31.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.31.mlp.shared_expert_gate.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.language_model.layers.31.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.31.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.language_model.layers.31.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.31.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.31.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.31.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.31.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.31.self_attn.o_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.language_model.layers.31.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.31.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.31.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.31.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.31.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_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.language_model.layers.32.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.32.input_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.language_model.layers.32.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.32.linear_attn.A_log" }, { "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.language_model.layers.32.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.32.linear_attn.conv1d.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.language_model.layers.32.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.32.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.32.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.32.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.32.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.32.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.32.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.32.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.32.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.32.linear_attn.in_proj_z.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.language_model.layers.32.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.32.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.32.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.32.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.32.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.32.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.32.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.32.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.32.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.32.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.32.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.32.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.32.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.32.mlp.shared_expert_gate.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.language_model.layers.32.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.32.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.language_model.layers.33.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.33.input_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.language_model.layers.33.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.33.linear_attn.A_log" }, { "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.language_model.layers.33.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.33.linear_attn.conv1d.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.language_model.layers.33.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.33.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.33.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.33.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.33.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.33.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.33.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.33.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.33.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.33.linear_attn.in_proj_z.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.language_model.layers.33.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.33.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.33.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.33.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.33.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.33.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.33.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.33.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.33.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.33.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.33.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.33.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.33.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.33.mlp.shared_expert_gate.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.language_model.layers.33.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.34.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.34.input_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.language_model.layers.34.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.34.linear_attn.A_log" }, { "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.language_model.layers.34.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.34.linear_attn.conv1d.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.language_model.layers.34.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.34.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.34.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.34.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.34.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.34.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.34.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.34.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.34.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.34.linear_attn.in_proj_z.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.language_model.layers.34.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.34.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.34.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.34.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.34.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.34.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.34.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.34.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.34.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.34.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.34.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.34.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.34.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.34.mlp.shared_expert_gate.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.language_model.layers.34.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.34.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.language_model.layers.35.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.35.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.35.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.35.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.35.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.35.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.35.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.35.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.35.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.35.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.35.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.35.mlp.shared_expert_gate.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.language_model.layers.35.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.35.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.language_model.layers.35.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.35.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.35.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.35.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.35.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.35.self_attn.o_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.language_model.layers.35.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.35.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.35.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.35.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.35.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_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.language_model.layers.36.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.36.input_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.language_model.layers.36.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.36.linear_attn.A_log" }, { "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.language_model.layers.36.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.36.linear_attn.conv1d.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.language_model.layers.36.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.36.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.36.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.36.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.36.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.36.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.36.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.36.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.36.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.36.linear_attn.in_proj_z.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.language_model.layers.36.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.36.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.36.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.36.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.36.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.36.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.36.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.36.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.36.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.36.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.36.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.36.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.36.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.36.mlp.shared_expert_gate.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.language_model.layers.36.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.36.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.language_model.layers.37.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.37.input_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.language_model.layers.37.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.37.linear_attn.A_log" }, { "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.language_model.layers.37.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.37.linear_attn.conv1d.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.language_model.layers.37.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.37.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.37.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.37.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.37.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.37.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.37.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.37.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.37.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.37.linear_attn.in_proj_z.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.language_model.layers.37.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.37.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.37.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.37.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.37.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.37.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.37.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.37.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.37.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.37.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.37.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.37.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.37.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.37.mlp.shared_expert_gate.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.language_model.layers.37.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_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.language_model.layers.38.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.38.input_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.language_model.layers.38.linear_attn.A_log", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.38.linear_attn.A_log" }, { "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.language_model.layers.38.linear_attn.conv1d", "outlier_strategy": "none", "parameters": 32768, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.38.linear_attn.conv1d.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.language_model.layers.38.linear_attn.dt_bias", "outlier_strategy": "none", "parameters": 32, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "attention", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.language_model.layers.38.linear_attn.dt_bias" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.38.linear_attn.in_proj_a", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.38.linear_attn.in_proj_a.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.38.linear_attn.in_proj_b", "outlier_strategy": "none", "parameters": 65536, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.38.linear_attn.in_proj_b.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.38.linear_attn.in_proj_qkv", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.38.linear_attn.in_proj_qkv.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.38.linear_attn.in_proj_z", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.38.linear_attn.in_proj_z.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.language_model.layers.38.linear_attn.norm", "outlier_strategy": "none", "parameters": 128, "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.language_model.layers.38.linear_attn.norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.38.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.38.linear_attn.out_proj.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.38.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.38.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.38.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.38.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.38.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.38.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.38.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.38.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.38.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.38.mlp.shared_expert_gate.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.language_model.layers.38.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.38.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.language_model.layers.39.input_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.39.input_layernorm.weight" }, { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.030937500000000003, "decode_latency_cost": 0.0, "hidden_state_error": 0.05500000000000001, "long_context_loss": 0.020625, "mtp_acceptance_loss": 0.0055000000000000005, "output_kl": 0.06875, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.044687500000000005, "token_disagreement": 0.037812500000000006 }, "module_path": "model.language_model.layers.39.mlp.gate", "outlier_strategy": "none", "parameters": 524288, "predicted_loss": 0.032456875, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.language_model.layers.39.mlp.gate.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.39.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.39.mlp.shared_expert.down_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.39.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.39.mlp.shared_expert.gate_proj.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.39.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.39.mlp.shared_expert.up_proj.weight" }, { "bits": 8, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.021876116042958818, "decode_latency_cost": 0.0, "hidden_state_error": 0.03889087296526012, "long_context_loss": 0.014584077361972545, "mtp_acceptance_loss": 0.003889087296526012, "output_kl": 0.04861359120657515, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.03159883428427385, "token_disagreement": 0.026737475163616336 }, "module_path": "model.language_model.layers.39.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.02295047640862413, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 9.0 }, "tensor": "model.language_model.layers.39.mlp.shared_expert_gate.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.language_model.layers.39.post_attention_layernorm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.layers.39.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.language_model.layers.39.self_attn.k_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.39.self_attn.k_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.39.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.39.self_attn.k_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.39.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.39.self_attn.o_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.language_model.layers.39.self_attn.q_norm", "outlier_strategy": "none", "parameters": 256, "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.language_model.layers.39.self_attn.q_norm.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.39.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_model.layers.39.self_attn.q_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.22910259710444142, "decode_latency_cost": 0.0, "hidden_state_error": 0.4072935059634514, "long_context_loss": 0.6363961030678927, "mtp_acceptance_loss": 0.04072935059634514, "output_kl": 0.5091168824543142, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.33092597359530423, "token_disagreement": 0.2800142853498728 }, "module_path": "model.language_model.layers.39.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.2645371321232617, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.language_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.language_model.norm", "outlier_strategy": "none", "parameters": 2048, "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.language_model.norm.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.visual.blocks.0.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.attn.proj.bias" }, { "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.visual.blocks.0.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.attn.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.visual.blocks.0.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.attn.qkv.bias" }, { "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.visual.blocks.0.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.attn.qkv.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.visual.blocks.0.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.mlp.linear_fc1.bias" }, { "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.visual.blocks.0.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.mlp.linear_fc1.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.visual.blocks.0.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.mlp.linear_fc2.bias" }, { "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.visual.blocks.0.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.mlp.linear_fc2.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.visual.blocks.0.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.norm1.bias" }, { "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.visual.blocks.0.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.norm1.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.visual.blocks.0.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.norm2.bias" }, { "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.visual.blocks.0.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.0.norm2.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.visual.blocks.1.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.attn.proj.bias" }, { "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.visual.blocks.1.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.attn.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.visual.blocks.1.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.attn.qkv.bias" }, { "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.visual.blocks.1.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.attn.qkv.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.visual.blocks.1.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.mlp.linear_fc1.bias" }, { "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.visual.blocks.1.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.mlp.linear_fc1.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.visual.blocks.1.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.mlp.linear_fc2.bias" }, { "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.visual.blocks.1.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.mlp.linear_fc2.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.visual.blocks.1.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.norm1.bias" }, { "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.visual.blocks.1.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.norm1.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.visual.blocks.1.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.norm2.bias" }, { "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.visual.blocks.1.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.1.norm2.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.visual.blocks.10.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.attn.proj.bias" }, { "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.visual.blocks.10.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.attn.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.visual.blocks.10.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.attn.qkv.bias" }, { "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.visual.blocks.10.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.attn.qkv.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.visual.blocks.10.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.mlp.linear_fc1.bias" }, { "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.visual.blocks.10.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.mlp.linear_fc1.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.visual.blocks.10.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.mlp.linear_fc2.bias" }, { "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.visual.blocks.10.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.mlp.linear_fc2.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.visual.blocks.10.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.norm1.bias" }, { "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.visual.blocks.10.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.norm1.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.visual.blocks.10.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.norm2.bias" }, { "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.visual.blocks.10.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.10.norm2.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.visual.blocks.11.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.attn.proj.bias" }, { "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.visual.blocks.11.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.attn.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.visual.blocks.11.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.attn.qkv.bias" }, { "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.visual.blocks.11.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.attn.qkv.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.visual.blocks.11.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.mlp.linear_fc1.bias" }, { "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.visual.blocks.11.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.mlp.linear_fc1.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.visual.blocks.11.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.mlp.linear_fc2.bias" }, { "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.visual.blocks.11.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.mlp.linear_fc2.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.visual.blocks.11.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.norm1.bias" }, { "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.visual.blocks.11.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.norm1.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.visual.blocks.11.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.norm2.bias" }, { "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.visual.blocks.11.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.11.norm2.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.visual.blocks.12.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.attn.proj.bias" }, { "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.visual.blocks.12.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.attn.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.visual.blocks.12.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.attn.qkv.bias" }, { "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.visual.blocks.12.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.attn.qkv.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.visual.blocks.12.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.mlp.linear_fc1.bias" }, { "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.visual.blocks.12.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.mlp.linear_fc1.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.visual.blocks.12.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.mlp.linear_fc2.bias" }, { "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.visual.blocks.12.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.mlp.linear_fc2.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.visual.blocks.12.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.norm1.bias" }, { "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.visual.blocks.12.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.norm1.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.visual.blocks.12.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.norm2.bias" }, { "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.visual.blocks.12.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.12.norm2.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.visual.blocks.13.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.attn.proj.bias" }, { "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.visual.blocks.13.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.attn.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.visual.blocks.13.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.attn.qkv.bias" }, { "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.visual.blocks.13.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.attn.qkv.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.visual.blocks.13.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.mlp.linear_fc1.bias" }, { "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.visual.blocks.13.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.mlp.linear_fc1.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.visual.blocks.13.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.mlp.linear_fc2.bias" }, { "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.visual.blocks.13.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.mlp.linear_fc2.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.visual.blocks.13.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.norm1.bias" }, { "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.visual.blocks.13.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.norm1.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.visual.blocks.13.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.norm2.bias" }, { "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.visual.blocks.13.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.13.norm2.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.visual.blocks.14.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.attn.proj.bias" }, { "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.visual.blocks.14.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.attn.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.visual.blocks.14.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.attn.qkv.bias" }, { "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.visual.blocks.14.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.attn.qkv.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.visual.blocks.14.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.mlp.linear_fc1.bias" }, { "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.visual.blocks.14.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.mlp.linear_fc1.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.visual.blocks.14.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.mlp.linear_fc2.bias" }, { "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.visual.blocks.14.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.mlp.linear_fc2.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.visual.blocks.14.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.norm1.bias" }, { "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.visual.blocks.14.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.norm1.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.visual.blocks.14.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.norm2.bias" }, { "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.visual.blocks.14.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.14.norm2.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.visual.blocks.15.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.attn.proj.bias" }, { "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.visual.blocks.15.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.attn.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.visual.blocks.15.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.attn.qkv.bias" }, { "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.visual.blocks.15.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.attn.qkv.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.visual.blocks.15.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.mlp.linear_fc1.bias" }, { "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.visual.blocks.15.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.mlp.linear_fc1.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.visual.blocks.15.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.mlp.linear_fc2.bias" }, { "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.visual.blocks.15.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.mlp.linear_fc2.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.visual.blocks.15.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.norm1.bias" }, { "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.visual.blocks.15.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.norm1.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.visual.blocks.15.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.norm2.bias" }, { "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.visual.blocks.15.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.15.norm2.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.visual.blocks.16.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.attn.proj.bias" }, { "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.visual.blocks.16.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.attn.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.visual.blocks.16.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.attn.qkv.bias" }, { "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.visual.blocks.16.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.attn.qkv.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.visual.blocks.16.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.mlp.linear_fc1.bias" }, { "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.visual.blocks.16.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.mlp.linear_fc1.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.visual.blocks.16.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.mlp.linear_fc2.bias" }, { "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.visual.blocks.16.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.mlp.linear_fc2.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.visual.blocks.16.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.norm1.bias" }, { "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.visual.blocks.16.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.norm1.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.visual.blocks.16.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.norm2.bias" }, { "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.visual.blocks.16.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.16.norm2.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.visual.blocks.17.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.attn.proj.bias" }, { "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.visual.blocks.17.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.attn.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.visual.blocks.17.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.attn.qkv.bias" }, { "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.visual.blocks.17.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.attn.qkv.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.visual.blocks.17.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.mlp.linear_fc1.bias" }, { "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.visual.blocks.17.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.mlp.linear_fc1.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.visual.blocks.17.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.mlp.linear_fc2.bias" }, { "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.visual.blocks.17.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.mlp.linear_fc2.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.visual.blocks.17.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.norm1.bias" }, { "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.visual.blocks.17.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.norm1.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.visual.blocks.17.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.norm2.bias" }, { "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.visual.blocks.17.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.17.norm2.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.visual.blocks.18.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.attn.proj.bias" }, { "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.visual.blocks.18.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.attn.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.visual.blocks.18.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.attn.qkv.bias" }, { "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.visual.blocks.18.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.attn.qkv.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.visual.blocks.18.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.mlp.linear_fc1.bias" }, { "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.visual.blocks.18.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.mlp.linear_fc1.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.visual.blocks.18.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.mlp.linear_fc2.bias" }, { "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.visual.blocks.18.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.mlp.linear_fc2.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.visual.blocks.18.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.norm1.bias" }, { "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.visual.blocks.18.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.norm1.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.visual.blocks.18.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.norm2.bias" }, { "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.visual.blocks.18.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.18.norm2.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.visual.blocks.19.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.attn.proj.bias" }, { "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.visual.blocks.19.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.attn.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.visual.blocks.19.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.attn.qkv.bias" }, { "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.visual.blocks.19.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.attn.qkv.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.visual.blocks.19.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.mlp.linear_fc1.bias" }, { "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.visual.blocks.19.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.mlp.linear_fc1.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.visual.blocks.19.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.mlp.linear_fc2.bias" }, { "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.visual.blocks.19.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.mlp.linear_fc2.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.visual.blocks.19.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.norm1.bias" }, { "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.visual.blocks.19.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.norm1.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.visual.blocks.19.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.norm2.bias" }, { "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.visual.blocks.19.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.19.norm2.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.visual.blocks.2.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.attn.proj.bias" }, { "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.visual.blocks.2.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.attn.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.visual.blocks.2.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.attn.qkv.bias" }, { "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.visual.blocks.2.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.attn.qkv.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.visual.blocks.2.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.mlp.linear_fc1.bias" }, { "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.visual.blocks.2.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.mlp.linear_fc1.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.visual.blocks.2.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.mlp.linear_fc2.bias" }, { "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.visual.blocks.2.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.mlp.linear_fc2.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.visual.blocks.2.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.norm1.bias" }, { "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.visual.blocks.2.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.norm1.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.visual.blocks.2.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.norm2.bias" }, { "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.visual.blocks.2.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.2.norm2.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.visual.blocks.20.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.attn.proj.bias" }, { "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.visual.blocks.20.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.attn.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.visual.blocks.20.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.attn.qkv.bias" }, { "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.visual.blocks.20.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.attn.qkv.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.visual.blocks.20.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.mlp.linear_fc1.bias" }, { "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.visual.blocks.20.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.mlp.linear_fc1.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.visual.blocks.20.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.mlp.linear_fc2.bias" }, { "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.visual.blocks.20.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.mlp.linear_fc2.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.visual.blocks.20.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.norm1.bias" }, { "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.visual.blocks.20.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.norm1.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.visual.blocks.20.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.norm2.bias" }, { "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.visual.blocks.20.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.20.norm2.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.visual.blocks.21.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.attn.proj.bias" }, { "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.visual.blocks.21.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.attn.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.visual.blocks.21.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.attn.qkv.bias" }, { "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.visual.blocks.21.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.attn.qkv.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.visual.blocks.21.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.mlp.linear_fc1.bias" }, { "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.visual.blocks.21.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.mlp.linear_fc1.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.visual.blocks.21.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.mlp.linear_fc2.bias" }, { "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.visual.blocks.21.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.mlp.linear_fc2.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.visual.blocks.21.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.norm1.bias" }, { "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.visual.blocks.21.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.norm1.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.visual.blocks.21.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.norm2.bias" }, { "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.visual.blocks.21.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.21.norm2.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.visual.blocks.22.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.attn.proj.bias" }, { "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.visual.blocks.22.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.attn.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.visual.blocks.22.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.attn.qkv.bias" }, { "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.visual.blocks.22.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.attn.qkv.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.visual.blocks.22.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.mlp.linear_fc1.bias" }, { "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.visual.blocks.22.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.mlp.linear_fc1.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.visual.blocks.22.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.mlp.linear_fc2.bias" }, { "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.visual.blocks.22.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.mlp.linear_fc2.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.visual.blocks.22.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.norm1.bias" }, { "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.visual.blocks.22.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.norm1.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.visual.blocks.22.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.norm2.bias" }, { "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.visual.blocks.22.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.22.norm2.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.visual.blocks.23.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.attn.proj.bias" }, { "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.visual.blocks.23.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.attn.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.visual.blocks.23.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.attn.qkv.bias" }, { "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.visual.blocks.23.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.attn.qkv.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.visual.blocks.23.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.mlp.linear_fc1.bias" }, { "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.visual.blocks.23.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.mlp.linear_fc1.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.visual.blocks.23.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.mlp.linear_fc2.bias" }, { "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.visual.blocks.23.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.mlp.linear_fc2.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.visual.blocks.23.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.norm1.bias" }, { "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.visual.blocks.23.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.norm1.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.visual.blocks.23.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.norm2.bias" }, { "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.visual.blocks.23.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.23.norm2.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.visual.blocks.24.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.attn.proj.bias" }, { "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.visual.blocks.24.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.attn.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.visual.blocks.24.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.attn.qkv.bias" }, { "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.visual.blocks.24.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.attn.qkv.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.visual.blocks.24.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.mlp.linear_fc1.bias" }, { "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.visual.blocks.24.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.mlp.linear_fc1.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.visual.blocks.24.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.mlp.linear_fc2.bias" }, { "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.visual.blocks.24.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.mlp.linear_fc2.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.visual.blocks.24.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.norm1.bias" }, { "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.visual.blocks.24.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.norm1.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.visual.blocks.24.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.norm2.bias" }, { "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.visual.blocks.24.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.24.norm2.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.visual.blocks.25.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.attn.proj.bias" }, { "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.visual.blocks.25.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.attn.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.visual.blocks.25.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.attn.qkv.bias" }, { "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.visual.blocks.25.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.attn.qkv.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.visual.blocks.25.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.mlp.linear_fc1.bias" }, { "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.visual.blocks.25.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.mlp.linear_fc1.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.visual.blocks.25.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.mlp.linear_fc2.bias" }, { "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.visual.blocks.25.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.mlp.linear_fc2.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.visual.blocks.25.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.norm1.bias" }, { "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.visual.blocks.25.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.norm1.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.visual.blocks.25.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.norm2.bias" }, { "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.visual.blocks.25.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.25.norm2.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.visual.blocks.26.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.attn.proj.bias" }, { "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.visual.blocks.26.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.attn.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.visual.blocks.26.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.attn.qkv.bias" }, { "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.visual.blocks.26.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.attn.qkv.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.visual.blocks.26.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.mlp.linear_fc1.bias" }, { "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.visual.blocks.26.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.mlp.linear_fc1.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.visual.blocks.26.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.mlp.linear_fc2.bias" }, { "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.visual.blocks.26.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.mlp.linear_fc2.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.visual.blocks.26.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.norm1.bias" }, { "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.visual.blocks.26.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.norm1.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.visual.blocks.26.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.norm2.bias" }, { "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.visual.blocks.26.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.26.norm2.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.visual.blocks.3.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.attn.proj.bias" }, { "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.visual.blocks.3.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.attn.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.visual.blocks.3.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.attn.qkv.bias" }, { "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.visual.blocks.3.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.attn.qkv.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.visual.blocks.3.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.mlp.linear_fc1.bias" }, { "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.visual.blocks.3.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.mlp.linear_fc1.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.visual.blocks.3.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.mlp.linear_fc2.bias" }, { "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.visual.blocks.3.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.mlp.linear_fc2.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.visual.blocks.3.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.norm1.bias" }, { "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.visual.blocks.3.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.norm1.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.visual.blocks.3.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.norm2.bias" }, { "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.visual.blocks.3.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.3.norm2.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.visual.blocks.4.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.attn.proj.bias" }, { "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.visual.blocks.4.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.attn.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.visual.blocks.4.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.attn.qkv.bias" }, { "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.visual.blocks.4.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.attn.qkv.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.visual.blocks.4.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.mlp.linear_fc1.bias" }, { "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.visual.blocks.4.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.mlp.linear_fc1.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.visual.blocks.4.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.mlp.linear_fc2.bias" }, { "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.visual.blocks.4.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.mlp.linear_fc2.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.visual.blocks.4.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.norm1.bias" }, { "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.visual.blocks.4.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.norm1.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.visual.blocks.4.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.norm2.bias" }, { "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.visual.blocks.4.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.4.norm2.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.visual.blocks.5.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.attn.proj.bias" }, { "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.visual.blocks.5.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.attn.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.visual.blocks.5.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.attn.qkv.bias" }, { "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.visual.blocks.5.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.attn.qkv.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.visual.blocks.5.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.mlp.linear_fc1.bias" }, { "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.visual.blocks.5.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.mlp.linear_fc1.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.visual.blocks.5.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.mlp.linear_fc2.bias" }, { "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.visual.blocks.5.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.mlp.linear_fc2.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.visual.blocks.5.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.norm1.bias" }, { "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.visual.blocks.5.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.norm1.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.visual.blocks.5.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.norm2.bias" }, { "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.visual.blocks.5.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.5.norm2.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.visual.blocks.6.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.attn.proj.bias" }, { "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.visual.blocks.6.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.attn.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.visual.blocks.6.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.attn.qkv.bias" }, { "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.visual.blocks.6.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.attn.qkv.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.visual.blocks.6.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.mlp.linear_fc1.bias" }, { "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.visual.blocks.6.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.mlp.linear_fc1.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.visual.blocks.6.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.mlp.linear_fc2.bias" }, { "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.visual.blocks.6.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.mlp.linear_fc2.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.visual.blocks.6.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.norm1.bias" }, { "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.visual.blocks.6.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.norm1.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.visual.blocks.6.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.norm2.bias" }, { "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.visual.blocks.6.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.6.norm2.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.visual.blocks.7.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.attn.proj.bias" }, { "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.visual.blocks.7.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.attn.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.visual.blocks.7.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.attn.qkv.bias" }, { "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.visual.blocks.7.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.attn.qkv.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.visual.blocks.7.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.mlp.linear_fc1.bias" }, { "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.visual.blocks.7.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.mlp.linear_fc1.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.visual.blocks.7.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.mlp.linear_fc2.bias" }, { "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.visual.blocks.7.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.mlp.linear_fc2.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.visual.blocks.7.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.norm1.bias" }, { "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.visual.blocks.7.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.norm1.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.visual.blocks.7.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.norm2.bias" }, { "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.visual.blocks.7.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.7.norm2.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.visual.blocks.8.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.attn.proj.bias" }, { "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.visual.blocks.8.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.attn.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.visual.blocks.8.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.attn.qkv.bias" }, { "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.visual.blocks.8.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.attn.qkv.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.visual.blocks.8.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.mlp.linear_fc1.bias" }, { "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.visual.blocks.8.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.mlp.linear_fc1.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.visual.blocks.8.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.mlp.linear_fc2.bias" }, { "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.visual.blocks.8.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.mlp.linear_fc2.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.visual.blocks.8.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.norm1.bias" }, { "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.visual.blocks.8.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.norm1.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.visual.blocks.8.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.norm2.bias" }, { "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.visual.blocks.8.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.8.norm2.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.visual.blocks.9.attn.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.attn.proj.bias" }, { "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.visual.blocks.9.attn.proj", "outlier_strategy": "none", "parameters": 1327104, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.attn.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.visual.blocks.9.attn.qkv.bias", "outlier_strategy": "none", "parameters": 3456, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.attn.qkv.bias" }, { "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.visual.blocks.9.attn.qkv", "outlier_strategy": "none", "parameters": 3981312, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.attn.qkv.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.visual.blocks.9.mlp.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4304, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.mlp.linear_fc1.bias" }, { "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.visual.blocks.9.mlp.linear_fc1", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.mlp.linear_fc1.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.visual.blocks.9.mlp.linear_fc2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.mlp.linear_fc2.bias" }, { "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.visual.blocks.9.mlp.linear_fc2", "outlier_strategy": "none", "parameters": 4958208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.mlp.linear_fc2.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.visual.blocks.9.norm1.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.norm1.bias" }, { "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.visual.blocks.9.norm1", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.norm1.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.visual.blocks.9.norm2.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.norm2.bias" }, { "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.visual.blocks.9.norm2", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.blocks.9.norm2.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.visual.merger.linear_fc1.bias", "outlier_strategy": "none", "parameters": 4608, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.merger.linear_fc1.bias" }, { "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.visual.merger.linear_fc1", "outlier_strategy": "none", "parameters": 21233664, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.merger.linear_fc1.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.visual.merger.linear_fc2.bias", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.merger.linear_fc2.bias" }, { "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.visual.merger.linear_fc2", "outlier_strategy": "none", "parameters": 9437184, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.merger.linear_fc2.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.visual.merger.norm.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.merger.norm.bias" }, { "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.visual.merger.norm", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.merger.norm.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.visual.patch_embed.proj.bias", "outlier_strategy": "none", "parameters": 1152, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.patch_embed.proj.bias" }, { "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.visual.patch_embed.proj", "outlier_strategy": "none", "parameters": 1769472, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.patch_embed.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.visual.pos_embed", "outlier_strategy": "none", "parameters": 2654208, "predicted_loss": 0.0, "reason": "protected vision policy", "role": "vision", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.visual.pos_embed.weight" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.0.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.0.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.0.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.1.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.1.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.1.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.10.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.10.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.10.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.11.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.11.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.11.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.12.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.12.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.12.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.13.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.13.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.13.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.14.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.14.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.14.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.15.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.15.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.15.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.16.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.16.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.16.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.17.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.17.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.17.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.18.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.18.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.18.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.19.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.19.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.19.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.2.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.2.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.2.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.20.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.20.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.20.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.21.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.21.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.21.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.22.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.22.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.22.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.23.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.23.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.23.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.24.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.24.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.24.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.25.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.25.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.25.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.26.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.26.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.26.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.26.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.27.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.27.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.27.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.27.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.28.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.28.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.28.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.28.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.29.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.29.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.29.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.29.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.3.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.3.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.3.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.30.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.30.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.30.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.30.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.31.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.31.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.31.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.31.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.32.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.32.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.32.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.32.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.33.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.33.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.33.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.33.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.34.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.34.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.34.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.34.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.35.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.35.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.35.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.35.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.36.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.36.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.36.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.36.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.37.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.37.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.37.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.37.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.38.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.38.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.38.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.38.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.39.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.39.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.39.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.39.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.4.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.4.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.4.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.5.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.5.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.5.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.6.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.6.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.6.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.7.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.7.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.7.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.8.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.8.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.8.mlp.experts.gate_up_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.9.mlp.experts.down_proj", "outlier_strategy": "none", "parameters": 268435456, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.mlp.experts.down_proj" }, { "bits": 4, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.171, "decode_latency_cost": 0.0, "hidden_state_error": 0.30400000000000005, "long_context_loss": 0.11399999999999999, "mtp_acceptance_loss": 0.0304, "output_kl": 0.38, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.24700000000000003, "token_disagreement": 0.20900000000000002 }, "module_path": "model.language_model.layers.9.mlp.experts.gate_up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.179398, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.language_model.layers.9.mlp.experts.gate_up_proj" } ], "calibration": null, "candidate_bits": [ 4, 6, 8, 16 ], "candidate_group_sizes": [ 32, 64 ], "constraints": { "effective_bpw_limit": 4.88, "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 }, "cost_model": "abstract-bpw", "created_at": "2026-08-13T21:15:46.911329Z", "effective_bpw": 4.879999995223769, "evidence_kind": "architecture_prior", "global_validation_required": true, "group_size": 32, "hardware": { "name": "ax-engine-apple-silicon-affine-dwq-v3", "runtime": "ax-engine", "supported_bits": [ 2, 3, 4, 6, 8, 16 ], "supported_group_sizes": [ 32, 64, 128 ], "supported_methods": [ "affine", "awq", "dwq", "gptq", "gptq-act", "bf16" ] }, "kernel_latency_host_id": null, "kernel_latency_sha256": null, "kv_cache": null, "method_near_ties": [], "method_near_ties_omitted": 0, "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": 4.387212400493483, "objective": { "cosine_distance": 0.03, "decode_latency_cost": 0.07, "hidden_state_error": 0.1, "long_context_loss": 0.05, "mtp_acceptance_loss": 0.22, "output_kl": 0.15, "peak_memory_cost": 0.03, "prefill_latency_cost": 0.03, "task_loss_delta": 0.2, "token_disagreement": 0.12 }, "primary_runtime": "ax-engine", "profile": "agent-coding", "quantizer": "axquant", "random_seed": 0, "schema_version": "axquant.plan.v1", "software_versions": { "ax_engine": null, "axquant": "1.6.2", "mlx": "0.32.0", "mlx_lm": "0.31.3", "pydantic": "2.13.4", "python": "3.13.14", "safetensors": "0.8.0" }, "source_model": { "architecture": "Qwen3_5MoeForConditionalGeneration", "format": "mlx", "local_path": null, "model_id": "deepreinforce-ai/Ornith-1.0-35B", "revision": "5df2ed3f675c7beaa490328cc70bb573b65fb660" }, "status": "planned", "target_bpw": 4.88, "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.", "target BPW raised from 4.0000 to 4.8800 to satisfy protection floors (policy minimum 4.8736)", "convert ladder: prior" ], "weight_distribution": { "4bit": { "fraction": 0.957675235263577, "parameters": 33621278720 }, "8bit": { "fraction": 0.015085597042949166, "parameters": 529612800 }, "bf16": { "fraction": 0.027239167693473853, "parameters": 956290416 } } }