{
  "analysis_sha256": "9c44404877f61bcdbcdc87bbb90214da353855153b72f0e4947a6f9051947e6d",
  "architecture_profile": {
    "adapter_id": "minicpm5-dense-v1",
    "config_model_type": "llama",
    "dense": true,
    "mtp_declared": false,
    "notes": [],
    "optimization_scope": "text-path",
    "product_family": "minicpm5",
    "support_level": "supported",
    "support_tier": "convertible",
    "text_layer_count": 24,
    "vision_present": false
  },
  "assignments": [
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "lm_head",
      "outlier_strategy": "none",
      "parameters": 200540160,
      "predicted_loss": 0.0,
      "reason": "protected lm_head policy",
      "role": "lm_head",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "lm_head.weight"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0253125,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.045000000000000005,
        "long_context_loss": 0.016875,
        "mtp_acceptance_loss": 0.0045000000000000005,
        "output_kl": 0.05625,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.036562500000000005,
        "token_disagreement": 0.030937500000000003
      },
      "module_path": "model.embed_tokens",
      "outlier_strategy": "none",
      "parameters": 200540160,
      "predicted_loss": 0.0307125,
      "reason": "protected embedding policy",
      "role": "embedding",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 3,
        "storage_bpw": 8.5
      },
      "tensor": "model.embed_tokens.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.0.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "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": "model.layers.0.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "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": "model.layers.0.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.0.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.0.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.0.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.0.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.0.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": "model.layers.0.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.0.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.0.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.0.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.1.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.1.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": "model.layers.1.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.1.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": "model.layers.1.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.1.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.1.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.1.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.1.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.1.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": "model.layers.1.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.1.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.1.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.1.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.10.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.10.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": "model.layers.10.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.10.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": "model.layers.10.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.10.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.10.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.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": "model.layers.10.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.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": "model.layers.10.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.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": "model.layers.10.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.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": "model.layers.11.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.11.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.11.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": "model.layers.11.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.11.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": "model.layers.11.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.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": "model.layers.11.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.11.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.11.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.11.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.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": "model.layers.11.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.11.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.11.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.11.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.12.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.12.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": "model.layers.12.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.12.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": "model.layers.12.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.12.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.12.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.12.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.12.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.12.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": "model.layers.12.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.12.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.12.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.12.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.13.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.13.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": "model.layers.13.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.13.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": "model.layers.13.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.13.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.13.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.13.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.13.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.13.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": "model.layers.13.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.13.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.13.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.13.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.14.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.14.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": "model.layers.14.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.14.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": "model.layers.14.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.14.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.14.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.14.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.14.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.14.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": "model.layers.14.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.14.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.14.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.14.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.15.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.15.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": "model.layers.15.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.15.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": "model.layers.15.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.15.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.15.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.15.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.15.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.15.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": "model.layers.15.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.15.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.15.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.15.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.16.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.16.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": "model.layers.16.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.16.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": "model.layers.16.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.16.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.16.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.16.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.16.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.16.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": "model.layers.16.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.16.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.16.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.16.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.17.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.17.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": "model.layers.17.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.17.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": "model.layers.17.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.17.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.17.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.17.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.17.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.17.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": "model.layers.17.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.17.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.17.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.17.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.18.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.18.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": "model.layers.18.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.18.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": "model.layers.18.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.18.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.18.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.18.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.18.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.18.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": "model.layers.18.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.18.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.18.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.18.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.19.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.19.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": "model.layers.19.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.19.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": "model.layers.19.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.19.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.19.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.19.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.19.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.19.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": "model.layers.19.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.19.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.19.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.19.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.2.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.2.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": "model.layers.2.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.2.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": "model.layers.2.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.2.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.2.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.2.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.2.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.2.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": "model.layers.2.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.2.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.2.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.2.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.20.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.20.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": "model.layers.20.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.20.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": "model.layers.20.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.20.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.20.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.20.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.20.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.20.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": "model.layers.20.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.20.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.20.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.20.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.21.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.21.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": "model.layers.21.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.21.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": "model.layers.21.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.21.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.21.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.21.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.21.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.21.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": "model.layers.21.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.21.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.21.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.21.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.22.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.22.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": "model.layers.22.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.22.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": "model.layers.22.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.22.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.22.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.22.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.22.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.22.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": "model.layers.22.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.22.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.22.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.22.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.23.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.23.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": "model.layers.23.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.23.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": "model.layers.23.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.23.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.23.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.23.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.23.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.23.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": "model.layers.23.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.23.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.23.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.23.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.3.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.3.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": "model.layers.3.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.3.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": "model.layers.3.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.3.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.3.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.3.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.3.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.3.self_attn.o_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.3.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.3.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.3.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.3.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.4.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.4.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": "model.layers.4.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.4.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": "model.layers.4.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.4.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.4.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.4.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.4.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.4.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": "model.layers.4.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.4.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.4.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.4.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.5.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.5.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": "model.layers.5.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.5.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": "model.layers.5.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.5.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.5.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.5.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.5.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.5.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": "model.layers.5.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.5.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.5.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.5.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.6.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.6.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": "model.layers.6.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.6.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": "model.layers.6.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.6.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.6.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.6.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.6.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.6.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": "model.layers.6.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.6.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.6.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.6.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.7.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.7.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": "model.layers.7.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.7.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": "model.layers.7.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.7.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.7.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.7.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.7.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.7.self_attn.o_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.7.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.7.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.7.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.7.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.8.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.8.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": "model.layers.8.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.8.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": "model.layers.8.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.8.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.8.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.8.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.8.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.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": "model.layers.8.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.8.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.8.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.8.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.input_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 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": "model.layers.9.mlp.down_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.9.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": "model.layers.9.mlp.gate_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.9.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": "model.layers.9.mlp.up_proj",
      "outlier_strategy": "none",
      "parameters": 7077888,
      "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": "model.layers.9.mlp.up_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.post_attention_layernorm.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.9.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.9.self_attn.k_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.324,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.576,
        "long_context_loss": 0.8999999999999999,
        "mtp_acceptance_loss": 0.0576,
        "output_kl": 0.72,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.46799999999999997,
        "token_disagreement": 0.396
      },
      "module_path": "model.layers.9.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.44783999999999996,
      "reason": "minimum storage candidate",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.9.self_attn.o_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.9.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 3145728,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.9.self_attn.q_proj.weight"
    },
    {
      "bits": 4,
      "group_size": 32,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.22910259710444142,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.4072935059634514,
        "long_context_loss": 0.6363961030678927,
        "mtp_acceptance_loss": 0.04072935059634514,
        "output_kl": 0.5091168824543142,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.33092597359530423,
        "token_disagreement": 0.2800142853498728
      },
      "module_path": "model.layers.9.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 393216,
      "predicted_loss": 0.31667070088658344,
      "reason": "selected by marginal quality gain per storage bit",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 7,
        "storage_bpw": 5.0
      },
      "tensor": "model.layers.9.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "hidden_state_error": 0.0,
        "long_context_loss": 0.0,
        "mtp_acceptance_loss": 0.0,
        "output_kl": 0.0,
        "peak_memory_cost": 0.0,
        "prefill_latency_cost": 0.0,
        "task_loss_delta": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.norm",
      "outlier_strategy": "none",
      "parameters": 1536,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.norm.weight"
    }
  ],
  "calibration": null,
  "candidate_bits": [
    4,
    6,
    8,
    16
  ],
  "candidate_group_sizes": [
    32,
    64
  ],
  "constraints": {
    "effective_bpw_limit": 7.38,
    "lm_head_min_bits": 16,
    "max_model_size_ratio_to_uniform4": 1.1,
    "minimum_mtp_acceptance_retention": 0.95,
    "minimum_mtp_speedup": 1.2,
    "minimum_quality_retention": 0.98
  },
  "created_at": "2026-08-02T18:28:16.773889Z",
  "effective_bpw": 7.379967222761561,
  "evidence_kind": "architecture_prior",
  "global_validation_required": true,
  "group_size": 32,
  "hardware": {
    "name": "ax-engine-apple-silicon-affine-dwq-v2",
    "runtime": "ax-engine",
    "supported_bits": [
      2,
      3,
      4,
      6,
      8,
      16
    ],
    "supported_group_sizes": [
      32,
      64,
      128
    ],
    "supported_methods": [
      "affine",
      "awq",
      "dwq",
      "bf16"
    ]
  },
  "kv_cache": null,
  "mtp": {
    "candidate_bits": [
      8,
      16
    ],
    "min_bits": 8,
    "mode": "protected",
    "optimize_for_acceptance": true,
    "preserve_external_sidecar": true,
    "protect_norms": true,
    "protect_output_head": true
  },
  "mtp_distribution": {},
  "nominal_bpw": 6.970061276094932,
  "objective": {
    "cosine_distance": 0.05,
    "decode_latency_cost": 0.05,
    "hidden_state_error": 0.12,
    "long_context_loss": 0.08,
    "mtp_acceptance_loss": 0.15,
    "output_kl": 0.25,
    "peak_memory_cost": 0.04,
    "prefill_latency_cost": 0.03,
    "task_loss_delta": 0.15,
    "token_disagreement": 0.08
  },
  "primary_runtime": "ax-engine",
  "profile": "general",
  "quantizer": "axquant",
  "random_seed": 0,
  "schema_version": "axquant.plan.v1",
  "software_versions": {
    "ax_engine": null,
    "axquant": "1.0.0",
    "mlx": "0.32.0",
    "mlx_lm": "0.31.3",
    "pydantic": "2.13.4",
    "python": "3.13.14",
    "safetensors": "0.8.0"
  },
  "source_model": {
    "architecture": "LlamaForCausalLM",
    "format": "mlx",
    "local_path": null,
    "model_id": "openbmb/MiniCPM5-1B",
    "revision": null
  },
  "status": "planned",
  "target_bpw": 7.38,
  "target_class": "4bit",
  "target_mode": "low-memory",
  "warnings": [
    "This report contains architecture priors, not calibration measurements.",
    "Conversion planning requires --allow-unmeasured for this report.",
    "Multi-group architecture priors weakly prefer smaller group sizes; this is development evidence only (AXQ-028).",
    "Plan uses non-release architecture_prior evidence and requires complete-model validation.",
    "target BPW raised from 6.0000 to 7.3800 to satisfy protection floors (policy minimum 7.3772)",
    "convert ladder: prior"
  ],
  "weight_distribution": {
    "4bit": {
      "fraction": 0.6287771645272388,
      "parameters": 679477248
    },
    "8bit": {
      "fraction": 0.18557659369727533,
      "parameters": 200540160
    },
    "bf16": {
      "fraction": 0.18564624177548586,
      "parameters": 200615424
    }
  }
}