{ "analysis_sha256": "e81d8aaff9e04bae7e7cc46d7ada02dc3c653334edddff68f53a690b1ace5eef", "architecture_profile": { "adapter_id": "qwen3-next-v1", "config_model_type": "qwen3_next", "dense": false, "mtp_declared": false, "notes": [ "Qwen3-Next hybrid MoE (e.g. Qwen3-Coder-Next): fused experts via MLX-LM.", "Development convert only; no coding-bench quality claim from architecture priors.", "MoE experts quantize as fused switch modules with uniform per-group precision; artifacts are development evidence until certified." ], "optimization_scope": "text-path", "product_family": "qwen3-next", "support_level": "supported", "support_tier": "convertible", "text_layer_count": 48, "vision_present": false }, "assignments": [ { "bits": 8, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.0253125, "decode_latency_cost": 0.0, "hidden_state_error": 0.045000000000000005, "long_context_loss": 0.016875, "mtp_acceptance_loss": 0.0045000000000000005, "output_kl": 0.05625, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.036562500000000005, "token_disagreement": 0.030937500000000003 }, "module_path": "model.embed_tokens", "outlier_strategy": "none", "parameters": 311164928, "predicted_loss": 0.0307125, "reason": "protected embedding policy", "role": "embedding", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.embed_tokens.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.0.input_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.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.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.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.layers.0.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.layers.0.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.0.linear_attn.in_proj_ba.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.layers.0.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.0.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.0.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.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.layers.0.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.0.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.0.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.0.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.0.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.0.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.0.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.0.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.0.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.0.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.0.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.0.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.0.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.0.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.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.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.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.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.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.layers.1.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.layers.1.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.1.linear_attn.in_proj_ba.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.layers.1.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.1.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.1.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.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.layers.1.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.1.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.1.mlp.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.layers.1.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.1.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.1.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.1.post_attention_layernorm.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.layers.1.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.1.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.1.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.1.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.1.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.1.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.1.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.1.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.1.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.2.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.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.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.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.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.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.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.layers.2.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.layers.2.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.2.linear_attn.in_proj_ba.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.layers.2.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.2.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.2.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.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.layers.2.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.2.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.2.mlp.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.layers.2.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.2.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.2.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.2.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.2.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.2.post_attention_layernorm.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.layers.2.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.2.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.2.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.2.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.2.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.2.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.2.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.3.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.3.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.3.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.3.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.3.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.3.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.3.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.3.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.3.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.3.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.3.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.3.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.3.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.3.post_attention_layernorm.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.layers.3.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.layers.3.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.3.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.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.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.layers.3.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.3.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.layers.3.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.3.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.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.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.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.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.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.layers.4.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.layers.4.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.4.linear_attn.in_proj_ba.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.layers.4.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.4.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.4.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.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.layers.4.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.4.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.4.mlp.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.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.layers.4.post_attention_layernorm.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.layers.4.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.4.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.4.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.4.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.4.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.4.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.4.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.4.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.4.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.4.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.4.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.5.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.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.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.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.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.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.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.layers.5.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.layers.5.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.5.linear_attn.in_proj_ba.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.layers.5.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.5.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.5.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.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.layers.5.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.5.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.5.mlp.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.layers.5.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.5.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.5.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.5.post_attention_layernorm.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.layers.5.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.5.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.5.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.5.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.5.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.5.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.5.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.5.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.5.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.6.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.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.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.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.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.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.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.layers.6.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.layers.6.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.6.linear_attn.in_proj_ba.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.layers.6.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.6.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.6.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.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.layers.6.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.6.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.6.mlp.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.layers.6.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.6.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.6.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.6.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.6.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.6.post_attention_layernorm.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.layers.6.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.6.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.6.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.6.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.6.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.6.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.6.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.7.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.7.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.7.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.7.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.7.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.7.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.7.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.7.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.7.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.7.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.7.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.7.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.7.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.7.post_attention_layernorm.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.layers.7.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.layers.7.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.7.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.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.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.layers.7.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.44783999999999996, "reason": "minimum storage candidate", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.7.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.layers.7.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.7.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.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.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.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.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.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.layers.8.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.layers.8.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.8.linear_attn.in_proj_ba.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.layers.8.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.8.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.8.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.layers.8.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.layers.8.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.8.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.8.mlp.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.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.layers.8.post_attention_layernorm.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.layers.8.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.8.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.8.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.8.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.8.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.8.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.8.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.8.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.8.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.8.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.8.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.9.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.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.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.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.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.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.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.layers.9.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.layers.9.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.9.linear_attn.in_proj_ba.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.layers.9.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.9.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.9.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.layers.9.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.layers.9.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.9.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.layers.9.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.9.mlp.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.layers.9.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.9.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.9.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.9.post_attention_layernorm.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.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.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.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.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.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.layers.10.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.layers.10.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.10.linear_attn.in_proj_ba.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.layers.10.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.10.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.10.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.layers.10.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.layers.10.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.10.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.10.mlp.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.layers.10.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.10.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.10.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.10.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.10.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.10.post_attention_layernorm.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.layers.9.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.9.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.9.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.9.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.9.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.9.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.9.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.9.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.9.mlp.switch_mlp.up_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.layers.10.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.10.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.10.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.10.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.10.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.10.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.10.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.11.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.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.layers.11.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.11.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.11.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.11.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.11.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.11.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.11.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.11.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.11.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.11.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.11.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.11.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.layers.11.post_attention_layernorm.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.layers.11.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.layers.11.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.11.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.layers.11.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.layers.11.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.layers.11.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.layers.11.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.layers.11.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.11.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.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.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.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.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.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.layers.12.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.layers.12.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.12.linear_attn.in_proj_ba.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.layers.12.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.12.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.12.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.layers.12.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.layers.12.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.12.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.12.mlp.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.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.layers.12.post_attention_layernorm.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.layers.12.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.12.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.12.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.12.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.12.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.12.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.12.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.12.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.12.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.12.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.12.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.13.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.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.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.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.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.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.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.layers.13.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.layers.13.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.13.linear_attn.in_proj_ba.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.layers.13.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.13.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.13.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.layers.13.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.layers.13.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.13.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.13.mlp.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.layers.13.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.13.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.13.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.13.post_attention_layernorm.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.layers.13.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.13.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.13.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.13.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.13.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.13.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.13.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.13.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.13.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.14.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.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.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.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.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.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.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.layers.14.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.layers.14.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.14.linear_attn.in_proj_ba.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.layers.14.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.14.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.14.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.layers.14.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.layers.14.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.14.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.14.mlp.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.layers.14.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.14.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.14.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.14.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.14.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.14.post_attention_layernorm.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.layers.14.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.14.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.14.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.14.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.14.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.14.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.14.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.15.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.15.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.15.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.15.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.15.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.15.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.15.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.15.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.15.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.15.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.15.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.15.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.15.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.layers.15.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.layers.15.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.15.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.layers.15.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.layers.15.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.layers.15.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.layers.15.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.layers.15.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.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.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.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.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.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.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.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.layers.16.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.layers.16.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.16.linear_attn.in_proj_ba.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.layers.16.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.16.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.16.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.layers.16.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.layers.16.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.16.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.16.mlp.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.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.layers.16.post_attention_layernorm.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.layers.16.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.16.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.16.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.16.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.16.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.16.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.16.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.16.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.16.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.16.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.16.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.17.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.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.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.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.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.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.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.layers.17.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.layers.17.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.17.linear_attn.in_proj_ba.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.layers.17.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.17.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.17.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.layers.17.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.layers.17.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.17.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.17.mlp.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.layers.17.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.17.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.17.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.17.post_attention_layernorm.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.layers.17.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.17.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.17.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.17.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.17.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.17.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.17.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.17.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.17.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.18.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.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.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.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.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.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.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.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.layers.18.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.18.linear_attn.in_proj_ba.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.layers.18.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.18.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.18.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.layers.18.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.layers.18.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.18.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.18.mlp.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.layers.18.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.18.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.18.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.18.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.18.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.18.post_attention_layernorm.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.layers.18.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.18.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.18.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.18.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.18.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.18.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.18.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.19.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.19.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.19.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.19.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.19.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.19.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.19.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.19.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.19.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.19.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.19.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.19.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.19.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.layers.19.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.layers.19.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.19.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.layers.19.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.layers.19.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.layers.19.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.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.layers.19.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.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.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.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.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.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.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.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.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.layers.20.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.20.linear_attn.in_proj_ba.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.layers.20.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.20.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.20.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.layers.20.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.layers.20.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.20.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.20.mlp.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.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.layers.20.post_attention_layernorm.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.layers.20.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.20.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.20.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.20.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.20.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.20.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.20.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.20.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.20.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.20.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.20.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.21.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.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.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.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.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.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.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.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.layers.21.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.21.linear_attn.in_proj_ba.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.layers.21.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.21.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.21.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.layers.21.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.layers.21.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.21.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.21.mlp.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.layers.21.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.21.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.21.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.21.post_attention_layernorm.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.layers.21.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.21.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.21.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.21.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.21.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.21.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.21.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.21.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.21.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.22.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.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.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.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.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.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.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.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.layers.22.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.22.linear_attn.in_proj_ba.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.layers.22.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.22.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.22.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.layers.22.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.layers.22.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.22.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.22.mlp.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.layers.22.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.22.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.22.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.22.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.22.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.22.post_attention_layernorm.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.layers.22.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.22.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.22.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.22.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.22.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.22.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.22.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.23.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.23.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.23.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.23.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.23.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.23.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.23.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.23.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.23.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.23.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.23.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.23.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.23.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.layers.23.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.23.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.layers.23.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.layers.23.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.layers.23.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.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.layers.23.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.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.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.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.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.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.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.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.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.layers.24.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.24.linear_attn.in_proj_ba.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.layers.24.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.24.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.24.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.layers.24.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.layers.24.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.24.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.24.mlp.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.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.layers.24.post_attention_layernorm.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.layers.24.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.24.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.24.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.24.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.24.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.24.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.24.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.24.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.24.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.24.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.24.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.25.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.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.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.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.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.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.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.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.layers.25.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.25.linear_attn.in_proj_ba.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.layers.25.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.25.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.25.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.layers.25.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.layers.25.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.25.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.25.mlp.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.layers.25.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.25.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.25.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.25.post_attention_layernorm.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.layers.25.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.25.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.25.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.25.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.25.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.25.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.25.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.25.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.25.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.26.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.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.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.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.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.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.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.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.layers.26.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.26.linear_attn.in_proj_ba.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.layers.26.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.26.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.26.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.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.layers.26.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.26.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.26.mlp.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.layers.26.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.26.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.26.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.26.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.26.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.26.post_attention_layernorm.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.layers.26.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.26.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.26.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.26.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.26.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.26.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.26.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.27.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.27.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.27.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.27.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.27.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.27.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.27.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.27.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.27.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.27.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.27.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.27.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.27.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.layers.27.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.27.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.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.layers.27.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.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.layers.27.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.27.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.28.input_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.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.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.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.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.layers.28.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.28.linear_attn.in_proj_ba.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.layers.28.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.28.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.28.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.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.layers.28.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.28.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.28.mlp.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.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.layers.28.post_attention_layernorm.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.layers.28.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.28.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.28.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.28.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.28.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.28.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.28.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.28.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.28.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.28.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.28.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.29.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.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.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.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.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.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.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.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.layers.29.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.29.linear_attn.in_proj_ba.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.layers.29.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.29.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.29.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.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.layers.29.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.29.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.29.mlp.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.layers.29.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.29.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.29.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.29.post_attention_layernorm.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.layers.29.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.29.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.29.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.29.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.29.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.29.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.29.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.29.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.29.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.30.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.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.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.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.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.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.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.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.layers.30.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.30.linear_attn.in_proj_ba.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.layers.30.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.30.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.30.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.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.layers.30.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.30.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.30.mlp.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.layers.30.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.30.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.30.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.30.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.30.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.30.post_attention_layernorm.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.layers.30.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.30.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.30.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.30.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.30.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.30.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.30.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.31.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.31.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.31.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.31.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.31.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.31.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.31.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.31.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.31.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.31.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.31.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.31.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.31.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.layers.31.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.31.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.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.layers.31.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.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.layers.31.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.31.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.32.input_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.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.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.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.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.layers.32.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.32.linear_attn.in_proj_ba.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.layers.32.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.32.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.32.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.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.layers.32.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.32.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.32.mlp.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.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.layers.32.post_attention_layernorm.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.layers.32.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.32.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.32.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.32.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.32.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.32.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.32.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.32.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.32.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.32.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.32.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.33.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.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.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.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.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.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.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.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.layers.33.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.33.linear_attn.in_proj_ba.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.layers.33.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.33.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.33.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.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.layers.33.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.33.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.33.mlp.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.layers.33.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.33.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.33.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.33.post_attention_layernorm.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.layers.33.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.33.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.33.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.33.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.33.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.33.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.33.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.33.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.33.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.34.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.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.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.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.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.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.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.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.layers.34.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.34.linear_attn.in_proj_ba.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.layers.34.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.34.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.34.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.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.layers.34.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.34.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.34.mlp.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.layers.34.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.34.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.34.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.34.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.34.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.34.post_attention_layernorm.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.layers.34.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.34.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.34.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.34.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.34.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.34.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.34.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.35.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.35.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.35.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.35.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.35.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.35.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.35.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.35.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.35.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.35.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.35.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.35.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.35.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.layers.35.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.35.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.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.layers.35.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.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.layers.35.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.35.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.36.input_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.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.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.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.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.layers.36.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.36.linear_attn.in_proj_ba.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.layers.36.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.36.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.36.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.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.layers.36.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.36.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.36.mlp.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.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.layers.36.post_attention_layernorm.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.layers.36.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.36.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.36.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.36.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.36.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.36.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.36.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.36.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.36.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.36.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.36.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.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.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.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.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.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.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.layers.37.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.37.linear_attn.in_proj_ba.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.layers.37.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.37.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.37.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.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.layers.37.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.37.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.37.mlp.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.layers.37.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.37.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.37.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.37.post_attention_layernorm.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.layers.37.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.37.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.37.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.37.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.37.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.37.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.37.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.37.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.37.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.38.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.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.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.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.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.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.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.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.layers.38.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.38.linear_attn.in_proj_ba.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.layers.38.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.38.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.38.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.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.layers.38.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.layers.38.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.38.mlp.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.layers.38.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.38.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.38.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.38.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.38.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.layers.38.post_attention_layernorm.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.layers.38.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.38.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.38.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.38.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.38.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.38.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.38.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.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.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.layers.39.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "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.layers.39.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.39.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "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.layers.39.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.39.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.39.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "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.layers.39.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.39.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.39.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.39.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.39.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.39.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.39.post_attention_layernorm", "outlier_strategy": "none", "parameters": 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.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.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.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.layers.39.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.39.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.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.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.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.layers.39.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.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.layers.39.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.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.layers.40.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.layers.40.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.layers.40.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.layers.40.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.layers.40.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.layers.40.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.layers.40.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.layers.40.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.layers.40.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.40.linear_attn.in_proj_ba.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.layers.40.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.40.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.40.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.layers.40.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.layers.40.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.40.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.layers.40.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.40.mlp.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.layers.40.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.layers.40.post_attention_layernorm.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.layers.40.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.40.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.layers.40.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.40.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.layers.40.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.40.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.40.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.layers.40.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.layers.40.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.40.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.40.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.40.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.40.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.40.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.41.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.layers.41.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.layers.41.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.layers.41.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.layers.41.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.layers.41.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.layers.41.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.layers.41.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.layers.41.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.41.linear_attn.in_proj_ba.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.layers.41.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.41.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.41.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.layers.41.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.layers.41.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.41.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.layers.41.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.41.mlp.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.layers.41.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.41.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.41.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.layers.41.post_attention_layernorm.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.layers.41.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.41.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.layers.41.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.41.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.layers.41.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.41.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.41.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.layers.41.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.layers.41.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.41.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.41.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.41.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.42.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.layers.42.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.layers.42.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.layers.42.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.layers.42.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.layers.42.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.layers.42.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.layers.42.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.layers.42.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.31667070088658344, "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.layers.42.linear_attn.in_proj_ba.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.layers.42.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.42.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.42.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.layers.42.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.layers.42.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.42.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.layers.42.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.42.mlp.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.layers.42.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.42.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.42.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.42.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.42.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.layers.42.post_attention_layernorm.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.layers.42.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.42.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.layers.42.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.42.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.layers.42.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.42.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.42.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.layers.42.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.layers.42.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.42.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.43.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.layers.43.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.layers.43.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.43.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.layers.43.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.43.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.layers.43.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.43.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.layers.43.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.43.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.43.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.layers.43.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.layers.43.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.43.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.43.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.43.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.43.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.43.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.43.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.layers.43.post_attention_layernorm.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.43.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.layers.43.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.layers.43.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.43.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.layers.43.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.43.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.layers.43.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.layers.43.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.layers.43.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.layers.43.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.layers.43.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.43.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.44.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.layers.44.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.layers.44.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.layers.44.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.layers.44.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.layers.44.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.layers.44.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.layers.44.linear_attn.dt_bias" }, { "bits": 6, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.081, "decode_latency_cost": 0.0, "hidden_state_error": 0.144, "long_context_loss": 0.22499999999999998, "mtp_acceptance_loss": 0.0144, "output_kl": 0.18, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.11699999999999999, "token_disagreement": 0.099 }, "module_path": "model.layers.44.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.11195999999999999, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 6.5 }, "tensor": "model.layers.44.linear_attn.in_proj_ba.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.layers.44.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.44.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.44.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.layers.44.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.layers.44.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.44.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.layers.44.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.44.mlp.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.layers.44.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.layers.44.post_attention_layernorm.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.layers.44.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.44.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.layers.44.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.44.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.layers.44.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.44.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.44.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.layers.44.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.layers.44.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.44.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.44.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.44.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.44.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.44.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.45.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.layers.45.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.layers.45.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.layers.45.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.layers.45.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.layers.45.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.layers.45.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.layers.45.linear_attn.dt_bias" }, { "bits": 6, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.081, "decode_latency_cost": 0.0, "hidden_state_error": 0.144, "long_context_loss": 0.22499999999999998, "mtp_acceptance_loss": 0.0144, "output_kl": 0.18, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.11699999999999999, "token_disagreement": 0.099 }, "module_path": "model.layers.45.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.11195999999999999, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 6.5 }, "tensor": "model.layers.45.linear_attn.in_proj_ba.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.layers.45.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.45.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.45.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.layers.45.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.layers.45.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.45.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.layers.45.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.45.mlp.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.layers.45.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.45.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.45.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.layers.45.post_attention_layernorm.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.layers.45.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.45.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.layers.45.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.45.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.layers.45.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.45.mlp.shared_expert.up_proj.weight" }, { "bits": 4, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.12091525958289964, "decode_latency_cost": 0.0, "hidden_state_error": 0.2149604614807105, "long_context_loss": 0.08061017305526641, "mtp_acceptance_loss": 0.021496046148071046, "output_kl": 0.2687005768508881, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.17465537495307726, "token_disagreement": 0.14778531726798846 }, "module_path": "model.layers.45.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.1467105149605849, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 5.0 }, "tensor": "model.layers.45.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.layers.45.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.45.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.45.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.45.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.46.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.layers.46.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.layers.46.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.layers.46.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.layers.46.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.layers.46.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.layers.46.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.layers.46.linear_attn.dt_bias" }, { "bits": 6, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.057275649276110355, "decode_latency_cost": 0.0, "hidden_state_error": 0.10182337649086284, "long_context_loss": 0.15909902576697318, "mtp_acceptance_loss": 0.010182337649086284, "output_kl": 0.12727922061357855, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.08273149339882606, "token_disagreement": 0.0700035713374682 }, "module_path": "model.layers.46.linear_attn.in_proj_ba", "outlier_strategy": "none", "parameters": 131072, "predicted_loss": 0.07916767522164586, "reason": "selected by marginal quality gain per storage bit", "role": "attention", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 7.0 }, "tensor": "model.layers.46.linear_attn.in_proj_ba.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.layers.46.linear_attn.in_proj_qkvz", "outlier_strategy": "none", "parameters": 25165824, "predicted_loss": 0.31667070088658344, "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.layers.46.linear_attn.in_proj_qkvz.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.46.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.layers.46.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.layers.46.linear_attn.out_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.46.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.layers.46.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.46.mlp.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.layers.46.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.46.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.46.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.46.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.46.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.layers.46.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": "lm_head", "outlier_strategy": "none", "parameters": 311164928, "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": 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.layers.46.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.46.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.layers.46.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.46.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.layers.46.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.46.mlp.shared_expert.up_proj.weight" }, { "bits": 6, "group_size": 64, "method": "affine", "metrics": { "cosine_distance": 0.04275, "decode_latency_cost": 0.0, "hidden_state_error": 0.07600000000000001, "long_context_loss": 0.028499999999999998, "mtp_acceptance_loss": 0.0076, "output_kl": 0.095, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.061750000000000006, "token_disagreement": 0.052250000000000005 }, "module_path": "model.layers.46.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.05187, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 6.5 }, "tensor": "model.layers.46.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.layers.46.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.46.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.47.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.layers.47.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.layers.47.mlp.gate", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.0375375, "reason": "protected router policy", "role": "router", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 3, "storage_bpw": 8.5 }, "tensor": "model.layers.47.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.layers.47.mlp.shared_expert.down_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.47.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.layers.47.mlp.shared_expert.gate_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.47.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.layers.47.mlp.shared_expert.up_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.20748, "reason": "minimum storage candidate", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 4.5 }, "tensor": "model.layers.47.mlp.shared_expert.up_proj.weight" }, { "bits": 6, "group_size": 32, "method": "affine", "metrics": { "cosine_distance": 0.03022881489572491, "decode_latency_cost": 0.0, "hidden_state_error": 0.05374011537017762, "long_context_loss": 0.020152543263816602, "mtp_acceptance_loss": 0.005374011537017762, "output_kl": 0.06717514421272203, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.043663843738269315, "token_disagreement": 0.036946329316997116 }, "module_path": "model.layers.47.mlp.shared_expert_gate", "outlier_strategy": "none", "parameters": 2048, "predicted_loss": 0.036677628740146226, "reason": "selected by marginal quality gain per storage bit", "role": "expert", "scale_strategy": "group-affine", "strategy_metadata": { "selected_from_candidates": 7, "storage_bpw": 7.0 }, "tensor": "model.layers.47.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.layers.47.mlp.switch_mlp.down_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.47.mlp.switch_mlp.down_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.47.mlp.switch_mlp.gate_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.47.mlp.switch_mlp.gate_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.47.mlp.switch_mlp.up_proj", "outlier_strategy": "none", "parameters": 536870912, "predicted_loss": 0.0, "reason": "non-quantizable tensor preserved", "role": "mlp", "scale_strategy": "none", "strategy_metadata": { "selected_from_candidates": 1, "storage_bpw": 16.0 }, "tensor": "model.layers.47.mlp.switch_mlp.up_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.47.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.layers.47.post_attention_layernorm.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.layers.47.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.layers.47.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.layers.47.self_attn.k_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.47.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.layers.47.self_attn.o_proj", "outlier_strategy": "none", "parameters": 8388608, "predicted_loss": 0.31667070088658344, "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.layers.47.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.layers.47.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.layers.47.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.layers.47.self_attn.q_proj", "outlier_strategy": "none", "parameters": 16777216, "predicted_loss": 0.31667070088658344, "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.layers.47.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.layers.47.self_attn.v_proj", "outlier_strategy": "none", "parameters": 1048576, "predicted_loss": 0.31667070088658344, "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.layers.47.self_attn.v_proj.weight" }, { "bits": 16, "group_size": null, "method": "bf16", "metrics": { "cosine_distance": 0.0, "decode_latency_cost": 0.0, "hidden_state_error": 0.0, "long_context_loss": 0.0, "mtp_acceptance_loss": 0.0, "output_kl": 0.0, "peak_memory_cost": 0.0, "prefill_latency_cost": 0.0, "task_loss_delta": 0.0, "token_disagreement": 0.0 }, "module_path": "model.norm", "outlier_strategy": "none", "parameters": 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.norm.weight" } ], "calibration": null, "candidate_bits": [ 4, 6, 8, 16 ], "candidate_group_sizes": [ 32, 64 ], "constraints": { "effective_bpw_limit": 15.73, "lm_head_min_bits": 16, "max_model_size_ratio_to_uniform4": 1.1, "minimum_mtp_acceptance_retention": 0.95, "minimum_mtp_speedup": 1.2, "minimum_quality_retention": 0.98 }, "created_at": "2026-08-02T21:28:18.222798Z", "effective_bpw": 15.729999988979143, "evidence_kind": "architecture_prior", "global_validation_required": true, "group_size": 32, "hardware": { "name": "ax-engine-apple-silicon-affine-dwq-v2", "runtime": "ax-engine", "supported_bits": [ 2, 3, 4, 6, 8, 16 ], "supported_group_sizes": [ 32, 64, 128 ], "supported_methods": [ "affine", "awq", "dwq", "bf16" ] }, "kv_cache": null, "mtp": { "candidate_bits": [ 8, 16 ], "min_bits": 8, "mode": "protected", "optimize_for_acceptance": true, "preserve_external_sidecar": true, "protect_norms": true, "protect_output_head": true }, "mtp_distribution": {}, "nominal_bpw": 15.709036958363761, "objective": { "cosine_distance": 0.05, "decode_latency_cost": 0.05, "hidden_state_error": 0.12, "long_context_loss": 0.08, "mtp_acceptance_loss": 0.15, "output_kl": 0.25, "peak_memory_cost": 0.04, "prefill_latency_cost": 0.03, "task_loss_delta": 0.15, "token_disagreement": 0.08 }, "primary_runtime": "ax-engine", "profile": "general", "quantizer": "axquant", "random_seed": 0, "schema_version": "axquant.plan.v1", "software_versions": { "ax_engine": null, "axquant": "1.0.1", "mlx": "0.32.0", "mlx_lm": "0.31.3", "pydantic": "2.13.4", "python": "3.13.14", "safetensors": "0.8.0" }, "source_model": { "architecture": "Qwen3NextForCausalLM", "format": "mlx", "local_path": null, "model_id": "Qwen/Qwen3-Coder-Next", "revision": null }, "status": "planned", "target_bpw": 15.73, "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 6.0000 to 15.7300 to satisfy protection floors (policy minimum 15.7219)", "convert ladder: prior" ], "weight_distribution": { "4bit": { "fraction": 0.021217981894828382, "parameters": 1690529792 }, "6bit": { "fraction": 4.9866963968879015e-06, "parameters": 397312 }, "8bit": { "fraction": 0.00453717399179112, "parameters": 361496576 }, "bf16": { "fraction": 0.9742398574169836, "parameters": 77621967616 } } }