{
  "analysis_sha256": "68715be29ca294eb05ca1e0ef2dd33f5b1b83ef4fc0977a5ff548e3d4497d8db",
  "architecture_profile": {
    "adapter_id": "deepseek-ocr2-v1",
    "audio_present": false,
    "config_model_type": "deepseekocr_2",
    "dense": false,
    "mtp_declared": false,
    "notes": [
      "DeepSeek-OCR-2 language MoE converts through MLX-VLM (deepseekocr_2).",
      "SAM / Qwen2 vision encoder / projector stay BF16-protected; OCR quality claims require measured eval — development evidence only.",
      "Prefer MLX BF16 sources (model_type=deepseekocr_2); official HF may need model_type remap from deepseek_vl_v2.",
      "MoE experts quantize as fused switch modules with uniform per-group precision; artifacts are development evidence until certified."
    ],
    "optimization_scope": "text-path",
    "product_family": "deepseek-ocr2",
    "support_level": "supported",
    "support_tier": "convertible",
    "text_layer_count": 12,
    "vision_present": true
  },
  "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": "language_model.model.embed_tokens",
      "outlier_strategy": "none",
      "parameters": 165478400,
      "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": "language_model.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": "language_model.model.layers.0.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.0.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.2025,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.36000000000000004,
        "long_context_loss": 0.135,
        "mtp_acceptance_loss": 0.036000000000000004,
        "output_kl": 0.45,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.29250000000000004,
        "token_disagreement": 0.24750000000000003
      },
      "module_path": "language_model.model.layers.0.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 8765440,
      "predicted_loss": 0.2457,
      "reason": "minimum storage candidate",
      "role": "mlp",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "language_model.model.layers.0.mlp.down_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.2025,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.36000000000000004,
        "long_context_loss": 0.135,
        "mtp_acceptance_loss": 0.036000000000000004,
        "output_kl": 0.45,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.29250000000000004,
        "token_disagreement": 0.24750000000000003
      },
      "module_path": "language_model.model.layers.0.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 8765440,
      "predicted_loss": 0.2457,
      "reason": "minimum storage candidate",
      "role": "mlp",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "language_model.model.layers.0.mlp.gate_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.2025,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.36000000000000004,
        "long_context_loss": 0.135,
        "mtp_acceptance_loss": 0.036000000000000004,
        "output_kl": 0.45,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.29250000000000004,
        "token_disagreement": 0.24750000000000003
      },
      "module_path": "language_model.model.layers.0.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 8765440,
      "predicted_loss": 0.2457,
      "reason": "minimum storage candidate",
      "role": "mlp",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "language_model.model.layers.0.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.model.layers.0.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.0.post_attention_layernorm.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": "language_model.model.layers.0.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.0.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": "language_model.model.layers.0.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.0.self_attn.o_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": "language_model.model.layers.0.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.0.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.0.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.0.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.model.layers.1.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.1.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.1.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.1.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.1.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.1.mlp.shared_experts.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": "language_model.model.layers.1.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.1.mlp.shared_experts.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": "language_model.model.layers.1.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.1.mlp.shared_experts.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": "language_model.model.layers.1.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.1.mlp.switch_mlp.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": "language_model.model.layers.1.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.1.mlp.switch_mlp.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": "language_model.model.layers.1.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.1.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.1.post_attention_layernorm.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": "language_model.model.layers.1.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.1.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": "language_model.model.layers.1.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.1.self_attn.o_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": "language_model.model.layers.1.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.1.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.1.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.1.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.model.layers.2.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.2.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.2.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.2.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.2.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.2.mlp.shared_experts.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": "language_model.model.layers.2.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.2.mlp.shared_experts.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": "language_model.model.layers.2.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.2.mlp.shared_experts.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": "language_model.model.layers.2.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.2.mlp.switch_mlp.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": "language_model.model.layers.2.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.2.mlp.switch_mlp.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": "language_model.model.layers.2.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.2.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.2.post_attention_layernorm.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": "language_model.model.layers.2.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.2.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": "language_model.model.layers.2.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.2.self_attn.o_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": "language_model.model.layers.2.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.2.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.2.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.2.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": "language_model.model.layers.3.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.3.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.3.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.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": "language_model.model.layers.3.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.3.mlp.shared_experts.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": "language_model.model.layers.3.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.3.mlp.shared_experts.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": "language_model.model.layers.3.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.3.mlp.shared_experts.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": "language_model.model.layers.3.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.3.mlp.switch_mlp.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": "language_model.model.layers.3.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.3.mlp.switch_mlp.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": "language_model.model.layers.3.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.3.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.3.post_attention_layernorm.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": "language_model.model.layers.3.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.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": "language_model.model.layers.3.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.3.self_attn.o_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": "language_model.model.layers.3.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.3.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.3.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.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": "language_model.model.layers.4.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.4.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.4.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.4.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.4.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.4.mlp.shared_experts.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": "language_model.model.layers.4.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.4.mlp.shared_experts.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": "language_model.model.layers.4.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.4.mlp.shared_experts.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": "language_model.model.layers.4.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.4.mlp.switch_mlp.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": "language_model.model.layers.4.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.4.mlp.switch_mlp.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": "language_model.model.layers.4.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.4.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.4.post_attention_layernorm.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": "language_model.model.layers.4.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.4.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": "language_model.model.layers.4.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.4.self_attn.o_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": "language_model.model.layers.4.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.4.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.4.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.4.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.model.layers.5.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.5.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.5.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.5.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.5.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.5.mlp.shared_experts.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": "language_model.model.layers.5.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.5.mlp.shared_experts.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": "language_model.model.layers.5.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.5.mlp.shared_experts.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": "language_model.model.layers.5.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.5.mlp.switch_mlp.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": "language_model.model.layers.5.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.5.mlp.switch_mlp.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": "language_model.model.layers.5.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.5.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.5.post_attention_layernorm.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": "language_model.model.layers.5.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.5.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": "language_model.model.layers.5.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.5.self_attn.o_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": "language_model.model.layers.5.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.5.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.5.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.5.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.model.layers.6.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.6.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.6.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.6.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.6.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.6.mlp.shared_experts.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": "language_model.model.layers.6.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.6.mlp.shared_experts.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": "language_model.model.layers.6.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.6.mlp.shared_experts.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": "language_model.model.layers.6.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.6.mlp.switch_mlp.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": "language_model.model.layers.6.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.6.mlp.switch_mlp.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": "language_model.model.layers.6.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.6.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.6.post_attention_layernorm.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": "language_model.model.layers.6.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.6.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": "language_model.model.layers.6.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.6.self_attn.o_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": "language_model.model.layers.6.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.6.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.6.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.6.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": "language_model.model.layers.7.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.7.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.7.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.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": "language_model.model.layers.7.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.7.mlp.shared_experts.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": "language_model.model.layers.7.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.7.mlp.shared_experts.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": "language_model.model.layers.7.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.7.mlp.shared_experts.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": "language_model.model.layers.7.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.7.mlp.switch_mlp.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": "language_model.model.layers.7.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.7.mlp.switch_mlp.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": "language_model.model.layers.7.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.7.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.7.post_attention_layernorm.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": "language_model.model.layers.7.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.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": "language_model.model.layers.7.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.7.self_attn.o_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": "language_model.model.layers.7.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.7.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.7.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.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": "language_model.model.layers.8.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.8.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.8.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.8.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.8.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.8.mlp.shared_experts.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": "language_model.model.layers.8.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.8.mlp.shared_experts.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": "language_model.model.layers.8.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.8.mlp.shared_experts.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": "language_model.model.layers.8.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.8.mlp.switch_mlp.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": "language_model.model.layers.8.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.8.mlp.switch_mlp.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": "language_model.model.layers.8.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.8.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.8.post_attention_layernorm.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": "language_model.model.layers.8.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.8.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": "language_model.model.layers.8.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.8.self_attn.o_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": "language_model.model.layers.8.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.8.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.8.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.8.self_attn.v_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": "language_model.model.layers.9.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.9.mlp.switch_mlp.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": "language_model.model.layers.9.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.9.mlp.switch_mlp.up_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": "language_model.model.layers.9.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.9.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": "language_model.model.layers.9.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.9.self_attn.o_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": "language_model.model.layers.9.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.9.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.9.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.9.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.0.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.0.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.1.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.1.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.10.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.10.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.11.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.11.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.2.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.2.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.3.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.3.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.4.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.4.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.5.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.5.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.6.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.6.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.7.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.7.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 8128,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.8.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.8.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.attn.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.attn.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.attn.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.attn.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.attn.qkv.bias",
      "outlier_strategy": "none",
      "parameters": 2304,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.attn.qkv.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.attn.qkv",
      "outlier_strategy": "none",
      "parameters": 1769472,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.attn.qkv.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.attn.rel_pos_h",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.attn.rel_pos_h"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.attn.rel_pos_w",
      "outlier_strategy": "none",
      "parameters": 1728,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.attn.rel_pos_w"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.mlp.lin1.bias",
      "outlier_strategy": "none",
      "parameters": 3072,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.mlp.lin1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.mlp.lin1",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.mlp.lin1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.mlp.lin2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.mlp.lin2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.mlp.lin2",
      "outlier_strategy": "none",
      "parameters": 2359296,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.mlp.lin2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.norm1.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.norm1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.norm1",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.norm1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.norm2.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.norm2.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.blocks.9.norm2",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.blocks.9.norm2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.neck.0",
      "outlier_strategy": "none",
      "parameters": 196608,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.neck.0.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.neck.1.bias",
      "outlier_strategy": "none",
      "parameters": 256,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.neck.1.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.neck.1",
      "outlier_strategy": "none",
      "parameters": 256,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.neck.1.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.neck.2",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.neck.2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.neck.3.bias",
      "outlier_strategy": "none",
      "parameters": 256,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.neck.3.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.neck.3",
      "outlier_strategy": "none",
      "parameters": 256,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.neck.3.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.net_2",
      "outlier_strategy": "none",
      "parameters": 1179648,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.net_2.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.net_3",
      "outlier_strategy": "none",
      "parameters": 4128768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.net_3.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.patch_embed.proj.bias",
      "outlier_strategy": "none",
      "parameters": 768,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.patch_embed.proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.patch_embed.proj",
      "outlier_strategy": "none",
      "parameters": 589824,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.patch_embed.proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "sam_model.pos_embed",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "sam_model.pos_embed"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.0.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.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": "vision_model.qwen2_encoder.layers.0.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.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": "vision_model.qwen2_encoder.layers.0.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.0.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.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": "vision_model.qwen2_encoder.layers.0.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.0.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.0.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.1.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.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": "vision_model.qwen2_encoder.layers.1.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.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": "vision_model.qwen2_encoder.layers.1.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.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": "vision_model.qwen2_encoder.layers.1.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.1.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.1.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.10.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.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": "vision_model.qwen2_encoder.layers.10.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.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": "vision_model.qwen2_encoder.layers.10.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.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": "vision_model.qwen2_encoder.layers.10.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.10.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.10.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": "vision_model.qwen2_encoder.layers.11.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.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": "vision_model.qwen2_encoder.layers.11.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.mlp.down_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.11.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.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": "vision_model.qwen2_encoder.layers.11.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.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": "vision_model.qwen2_encoder.layers.11.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.11.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.11.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.11.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.11.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.11.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.11.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.11.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.11.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.12.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.12.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.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": "vision_model.qwen2_encoder.layers.12.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.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": "vision_model.qwen2_encoder.layers.12.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.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": "vision_model.qwen2_encoder.layers.12.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.12.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.12.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.13.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.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": "vision_model.qwen2_encoder.layers.13.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.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": "vision_model.qwen2_encoder.layers.13.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.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": "vision_model.qwen2_encoder.layers.13.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.13.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.13.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.14.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.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": "vision_model.qwen2_encoder.layers.14.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.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": "vision_model.qwen2_encoder.layers.14.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.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": "vision_model.qwen2_encoder.layers.14.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.14.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.14.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.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": "vision_model.qwen2_encoder.layers.15.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.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": "vision_model.qwen2_encoder.layers.15.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.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": "vision_model.qwen2_encoder.layers.15.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.15.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.15.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.15.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.15.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.16.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.16.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.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": "vision_model.qwen2_encoder.layers.16.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.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": "vision_model.qwen2_encoder.layers.16.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.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": "vision_model.qwen2_encoder.layers.16.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.16.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.16.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.17.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.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": "vision_model.qwen2_encoder.layers.17.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.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": "vision_model.qwen2_encoder.layers.17.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.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": "vision_model.qwen2_encoder.layers.17.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.17.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.17.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.18.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.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": "vision_model.qwen2_encoder.layers.18.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.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": "vision_model.qwen2_encoder.layers.18.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.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": "vision_model.qwen2_encoder.layers.18.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.18.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.18.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.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": "vision_model.qwen2_encoder.layers.19.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.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": "vision_model.qwen2_encoder.layers.19.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.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": "vision_model.qwen2_encoder.layers.19.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.19.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.19.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.19.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.19.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.2.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.2.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.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": "vision_model.qwen2_encoder.layers.2.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.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": "vision_model.qwen2_encoder.layers.2.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.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": "vision_model.qwen2_encoder.layers.2.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.2.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.2.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": "vision_model.qwen2_encoder.layers.20.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.20.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.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": "vision_model.qwen2_encoder.layers.20.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.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": "vision_model.qwen2_encoder.layers.20.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.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": "vision_model.qwen2_encoder.layers.20.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.20.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.20.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.21.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.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": "vision_model.qwen2_encoder.layers.21.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.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": "vision_model.qwen2_encoder.layers.21.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.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": "vision_model.qwen2_encoder.layers.21.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.21.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.21.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.22.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.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": "vision_model.qwen2_encoder.layers.22.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.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": "vision_model.qwen2_encoder.layers.22.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.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": "vision_model.qwen2_encoder.layers.22.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.22.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.22.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.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": "vision_model.qwen2_encoder.layers.23.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.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": "vision_model.qwen2_encoder.layers.23.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.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": "vision_model.qwen2_encoder.layers.23.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.23.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.23.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.23.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.23.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.3.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.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": "vision_model.qwen2_encoder.layers.3.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.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": "vision_model.qwen2_encoder.layers.3.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.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": "vision_model.qwen2_encoder.layers.3.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.3.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.3.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.3.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.3.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.3.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.3.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.3.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.3.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.3.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.4.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.4.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.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": "vision_model.qwen2_encoder.layers.4.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.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": "vision_model.qwen2_encoder.layers.4.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.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": "vision_model.qwen2_encoder.layers.4.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.4.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.4.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.5.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.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": "vision_model.qwen2_encoder.layers.5.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.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": "vision_model.qwen2_encoder.layers.5.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.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": "vision_model.qwen2_encoder.layers.5.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.5.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.5.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.6.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.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": "vision_model.qwen2_encoder.layers.6.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.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": "vision_model.qwen2_encoder.layers.6.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.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": "vision_model.qwen2_encoder.layers.6.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.6.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.6.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": "vision_model.qwen2_encoder.layers.7.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.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": "vision_model.qwen2_encoder.layers.7.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.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": "vision_model.qwen2_encoder.layers.7.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.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": "vision_model.qwen2_encoder.layers.7.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.7.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.7.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.7.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.7.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.7.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.7.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.7.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.7.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.7.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.8.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.8.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.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": "vision_model.qwen2_encoder.layers.8.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.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": "vision_model.qwen2_encoder.layers.8.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.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": "vision_model.qwen2_encoder.layers.8.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.8.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.8.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.9.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.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": "vision_model.qwen2_encoder.layers.9.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.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": "vision_model.qwen2_encoder.layers.9.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 4358144,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.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": "vision_model.qwen2_encoder.layers.9.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.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": "vision_model.qwen2_encoder.layers.9.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 802816,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 128,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.layers.9.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 114688,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.layers.9.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.norm",
      "outlier_strategy": "none",
      "parameters": 896,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.norm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.query_1024",
      "outlier_strategy": "none",
      "parameters": 229376,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.query_1024"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "vision_model.qwen2_encoder.query_768",
      "outlier_strategy": "none",
      "parameters": 129024,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "vision_model.qwen2_encoder.query_768"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.lm_head",
      "outlier_strategy": "none",
      "parameters": 165478400,
      "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": "language_model.lm_head.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "language_model.model.layers.10.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.10.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.10.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.10.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.10.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.10.mlp.shared_experts.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": "language_model.model.layers.10.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.10.mlp.shared_experts.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": "language_model.model.layers.10.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.10.mlp.shared_experts.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": "language_model.model.layers.10.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.10.mlp.switch_mlp.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": "language_model.model.layers.10.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.10.mlp.switch_mlp.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": "language_model.model.layers.10.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.10.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.10.post_attention_layernorm.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": "language_model.model.layers.10.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.10.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": "language_model.model.layers.10.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.10.self_attn.o_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": "language_model.model.layers.10.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.10.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.10.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.10.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": "language_model.model.layers.11.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.11.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.11.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.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": "language_model.model.layers.11.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.11.mlp.shared_experts.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": "language_model.model.layers.11.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.11.mlp.shared_experts.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": "language_model.model.layers.11.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.11.mlp.shared_experts.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": "language_model.model.layers.11.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.11.mlp.switch_mlp.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": "language_model.model.layers.11.mlp.switch_mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.model.layers.11.mlp.switch_mlp.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": "language_model.model.layers.11.mlp.switch_mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.11.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.11.post_attention_layernorm.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": "language_model.model.layers.11.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.11.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.11.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.11.self_attn.o_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": "language_model.model.layers.11.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.model.layers.11.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "language_model.model.layers.11.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1638400,
      "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": "language_model.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": "language_model.model.layers.9.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.layers.9.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.021876116042958818,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.03889087296526012,
        "long_context_loss": 0.014584077361972545,
        "mtp_acceptance_loss": 0.003889087296526012,
        "output_kl": 0.04861359120657515,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.03159883428427385,
        "token_disagreement": 0.026737475163616336
      },
      "module_path": "language_model.model.layers.9.mlp.gate",
      "outlier_strategy": "none",
      "parameters": 81920,
      "predicted_loss": 0.02654302079879003,
      "reason": "protected router policy",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 9.0
      },
      "tensor": "language_model.model.layers.9.mlp.gate.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.171,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.30400000000000005,
        "long_context_loss": 0.11399999999999999,
        "mtp_acceptance_loss": 0.0304,
        "output_kl": 0.38,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.24700000000000003,
        "token_disagreement": 0.20900000000000002
      },
      "module_path": "language_model.model.layers.9.mlp.shared_experts.down_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.9.mlp.shared_experts.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": "language_model.model.layers.9.mlp.shared_experts.gate_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.9.mlp.shared_experts.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": "language_model.model.layers.9.mlp.shared_experts.up_proj",
      "outlier_strategy": "none",
      "parameters": 2293760,
      "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": "language_model.model.layers.9.mlp.shared_experts.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": "language_model.model.layers.9.mlp.switch_mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 73400320,
      "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": "language_model.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": "language_model.model.layers.9.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.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": "language_model.model.norm",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "language_model.model.norm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "projector.layers.bias",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "projector.layers.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "projector.layers",
      "outlier_strategy": "none",
      "parameters": 1146880,
      "predicted_loss": 0.0,
      "reason": "protected vision policy",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "projector.layers.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "view_separator",
      "outlier_strategy": "none",
      "parameters": 1280,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "vision",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "view_separator"
    }
  ],
  "calibration": null,
  "candidate_bits": [
    4,
    6,
    8,
    16
  ],
  "candidate_group_sizes": [
    32,
    64
  ],
  "constraints": {
    "effective_bpw_limit": 6.8,
    "lm_head_min_bits": 16,
    "max_model_size_ratio_to_uniform4": 1.1,
    "minimum_mtp_acceptance_retention": 0.95,
    "minimum_mtp_speedup": 1.2,
    "minimum_quality_retention": 0.98
  },
  "cost_model": "abstract-bpw",
  "created_at": "2026-08-13T23:52:21.977663Z",
  "effective_bpw": 6.799939344715201,
  "evidence_kind": "architecture_prior",
  "global_validation_required": true,
  "group_size": 32,
  "hardware": {
    "name": "ax-engine-apple-silicon-affine-dwq-v3",
    "runtime": "ax-engine",
    "supported_bits": [
      2,
      3,
      4,
      6,
      8,
      16
    ],
    "supported_group_sizes": [
      32,
      64,
      128
    ],
    "supported_methods": [
      "affine",
      "awq",
      "dwq",
      "gptq",
      "gptq-act",
      "bf16"
    ]
  },
  "kernel_latency_host_id": null,
  "kernel_latency_sha256": null,
  "kv_cache": null,
  "method_near_ties": [],
  "method_near_ties_omitted": 0,
  "mtp": {
    "candidate_bits": [
      8,
      16
    ],
    "min_bits": 8,
    "mode": "protected",
    "optimize_for_acceptance": true,
    "preserve_external_sidecar": true,
    "protect_norms": true,
    "protect_output_head": true
  },
  "mtp_distribution": {},
  "nominal_bpw": 6.391260200407931,
  "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": "1.6.2",
    "mlx": "0.32.0",
    "mlx_lm": "0.31.3",
    "pydantic": "2.13.4",
    "python": "3.13.14",
    "safetensors": "0.8.0"
  },
  "source_model": {
    "architecture": "DeepseekOCR2ForCausalLM",
    "format": "mlx",
    "local_path": "/Volumes/Ext4T/axquant/work/deepseek-ocr2-axq-dev/src-deepseek-ocr2-bf16",
    "model_id": "deepseek-ai/DeepSeek-OCR-2",
    "revision": "aaa02f3811945a91062062994c5c4a3f4c0af2b0"
  },
  "status": "planned",
  "target_bpw": 6.8,
  "target_class": "6p8bpw",
  "target_mode": "low-memory",
  "warnings": [
    "This report contains architecture priors, not calibration measurements.",
    "Conversion planning requires --allow-unmeasured for this report.",
    "Multi-group architecture priors weakly prefer smaller group sizes; this is development evidence only (AXQ-028).",
    "Plan uses non-release architecture_prior evidence and requires complete-model validation.",
    "target BPW raised from 4.0000 to 6.8000 to satisfy protection floors (policy minimum 6.7998)",
    "convert ladder: prior",
    "kept 11 MoEGate router modules at BF16 (not nn.Linear; MLX-VLM quantize_model cannot pack them)"
  ],
  "weight_distribution": {
    "4bit": {
      "fraction": 0.7680001450288254,
      "parameters": 2602844160
    },
    "8bit": {
      "fraction": 0.049092257405770445,
      "parameters": 166379520
    },
    "bf16": {
      "fraction": 0.18290759756540412,
      "parameters": 619895680
    }
  }
}