{
  "note": "Hermes-tuned KXL policies for stock llama.cpp targeting Unsloth-style Dynamic XL buckets. These are not a claim of byte-identical Unsloth Dynamic 2.0 quants unless generated by Unsloth tooling.",
  "bf16_parent": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf",
  "imatrix": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf",
  "commands": [
    {
      "name": "Q8KXL",
      "base_quant": "Q8_0",
      "target_gb": 9.34,
      "tensor_type_file": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q8KXL.tensor-types.txt",
      "output": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q8KXL.gguf",
      "command": [
        "llama-quantize",
        "--tensor-type-file",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q8KXL.tensor-types.txt",
        "--imatrix",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q8KXL.gguf",
        "Q8_0"
      ],
      "policy": {
        "base": "Q8_0",
        "target_gb": 9.34,
        "rules": [
          {
            "match": "token_embd",
            "type": "F16",
            "reason": "embedding/tool-token sensitivity"
          },
          {
            "match": "output",
            "type": "F16",
            "reason": "logit head sensitivity"
          },
          {
            "match": "attn",
            "type": "Q8_0",
            "reason": "attention path retained high precision"
          },
          {
            "match": "gate",
            "type": "Q8_0",
            "reason": "router/gate routing sensitivity"
          },
          {
            "match": "router",
            "type": "Q8_0",
            "reason": "router sensitivity"
          },
          {
            "match": "norm",
            "type": "F16",
            "reason": "normalization stability"
          },
          {
            "match": "ffn",
            "type": "Q8_0",
            "reason": "expert/base MLP storage"
          }
        ]
      }
    },
    {
      "name": "Q6KXL",
      "base_quant": "Q6_K",
      "target_gb": 7.74,
      "tensor_type_file": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q6KXL.tensor-types.txt",
      "output": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q6KXL.gguf",
      "command": [
        "llama-quantize",
        "--tensor-type-file",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q6KXL.tensor-types.txt",
        "--imatrix",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q6KXL.gguf",
        "Q6_K"
      ],
      "policy": {
        "base": "Q6_K",
        "target_gb": 7.74,
        "rules": [
          {
            "match": "token_embd",
            "type": "Q8_0",
            "reason": "embedding/tool-token sensitivity"
          },
          {
            "match": "output",
            "type": "Q8_0",
            "reason": "logit head sensitivity"
          },
          {
            "match": "attn_q",
            "type": "Q8_0",
            "reason": "tool-token attention sensitivity"
          },
          {
            "match": "attn_k",
            "type": "Q8_0",
            "reason": "tool-token attention sensitivity"
          },
          {
            "match": "attn_v",
            "type": "Q8_0",
            "reason": "tool-token attention sensitivity"
          },
          {
            "match": "attn_output",
            "type": "Q8_0",
            "reason": "attention output sensitivity"
          },
          {
            "match": "gate",
            "type": "Q8_0",
            "reason": "router/gate routing sensitivity"
          },
          {
            "match": "router",
            "type": "Q8_0",
            "reason": "router sensitivity"
          },
          {
            "match": "norm",
            "type": "F16",
            "reason": "normalization stability"
          },
          {
            "match": "ffn",
            "type": "Q6_K",
            "reason": "default expert/base MLP target"
          }
        ]
      }
    },
    {
      "name": "Q5KXL",
      "base_quant": "Q5_K_M",
      "target_gb": 6.39,
      "tensor_type_file": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q5KXL.tensor-types.txt",
      "output": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q5KXL.gguf",
      "command": [
        "llama-quantize",
        "--tensor-type-file",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q5KXL.tensor-types.txt",
        "--imatrix",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q5KXL.gguf",
        "Q5_K_M"
      ],
      "policy": {
        "base": "Q5_K_M",
        "target_gb": 6.39,
        "rules": [
          {
            "match": "token_embd",
            "type": "Q8_0",
            "reason": "embedding/tool-token sensitivity"
          },
          {
            "match": "output",
            "type": "Q8_0",
            "reason": "logit head sensitivity"
          },
          {
            "match": "attn",
            "type": "Q6_K",
            "reason": "attention kept above base"
          },
          {
            "match": "gate",
            "type": "Q8_0",
            "reason": "router/gate routing sensitivity"
          },
          {
            "match": "router",
            "type": "Q8_0",
            "reason": "router sensitivity"
          },
          {
            "match": "norm",
            "type": "F16",
            "reason": "normalization stability"
          },
          {
            "match": "ffn",
            "type": "Q5_K_M",
            "reason": "default expert/base MLP target"
          }
        ]
      }
    },
    {
      "name": "Q4KXL",
      "base_quant": "Q4_K_M",
      "target_gb": 5.35,
      "tensor_type_file": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q4KXL.tensor-types.txt",
      "output": "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q4KXL.gguf",
      "command": [
        "llama-quantize",
        "--tensor-type-file",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/Q4KXL.tensor-types.txt",
        "--imatrix",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf",
        "release_work/lfm25-moe8-mac-trainer/artifacts/gguf/xl_quant_policies/LFM-2.5-8B-1B-Hermes-Tuned-Q4KXL.gguf",
        "Q4_K_M"
      ],
      "policy": {
        "base": "Q4_K_M",
        "target_gb": 5.35,
        "rules": [
          {
            "match": "token_embd",
            "type": "Q6_K",
            "reason": "embedding/tool-token sensitivity"
          },
          {
            "match": "output",
            "type": "Q6_K",
            "reason": "logit head sensitivity"
          },
          {
            "match": "attn_q",
            "type": "Q6_K",
            "reason": "attention query sensitivity"
          },
          {
            "match": "attn_k",
            "type": "Q6_K",
            "reason": "attention key sensitivity"
          },
          {
            "match": "attn_v",
            "type": "Q5_K_M",
            "reason": "attention value above base"
          },
          {
            "match": "attn_output",
            "type": "Q5_K_M",
            "reason": "attention output above base"
          },
          {
            "match": "gate",
            "type": "Q8_0",
            "reason": "router/gate routing sensitivity"
          },
          {
            "match": "router",
            "type": "Q8_0",
            "reason": "router sensitivity"
          },
          {
            "match": "norm",
            "type": "F16",
            "reason": "normalization stability"
          },
          {
            "match": "ffn",
            "type": "Q4_K_M",
            "reason": "default expert/base MLP target"
          }
        ]
      }
    }
  ]
}