{
  "schema": 2,
  "created_utc": "2026-08-25T20:43:40.846999+00:00",
  "mother": {
    "path": "SOURCE-MANIFEST-BF16-MOTHER.json",
    "sha256": "fed14fbd38a5b6369cf5dc201c6bba3bf4b70ea3887100e0737be4a57fc4a68a",
    "repo": "huihui-ai/Huihui-Qwen3.8-27B-abliterated",
    "resolved_sha": "739e3c5b89849f6c238ce1e5b70008612ae42cdd",
    "file_count": 32,
    "indexed_tensor_count": 1199,
    "indexed_total_size": 55562855904.0
  },
  "tool": {
    "name": "GPTQModel",
    "version": "7.3.2",
    "commit": "f6f1ab12fe8115f84c52253871ca039ac6c425ee",
    "base_py_sha256": "5550ed2e108e3513a5e7a39748c2d2d42c2b9d78ef3ebea0e7c37657ae8747bd"
  },
  "quantization_script_sha256": "f8a469ccbf47e5689b8fabcd38d1aafb4bd0de594d2b4ba6e4625f4f07c7ae29",
  "hybrid_smoothing_script_sha256": "4e540a56695d225fd8dc164f6b03af59190ef20d403d153cdf2338077790e035",
  "calibration_corpus_sha256": "de8c3c856c55afd065ade04929659cf990cde7b8cf5fd846c6ee1867b3a13690",
  "calibration_records": 320,
  "quantization_config": {
    "bits": 4,
    "checkpoint_format": "gemm",
    "desc_act": false,
    "dynamic": {
      "-:.*\\.linear_attn\\.(?:in_proj_a|in_proj_b)$": {},
      "-:.*\\.lm_head$": {},
      "-:.*\\.mtp\\..*": {},
      "-:.*\\.visual\\..*": {}
    },
    "format": "gemm",
    "group_size": 128,
    "lm_head": false,
    "meta": {
      "act_group_aware": null,
      "auto_forward_data_parallel": false,
      "calibration_corpus_sha256": "de8c3c856c55afd065ade04929659cf990cde7b8cf5fd846c6ee1867b3a13690",
      "calibration_record_count": 320,
      "damp_auto_increment": null,
      "damp_percent": null,
      "dense_vram_strategy": "exclusive",
      "dense_vram_strategy_devices": [
        "cuda:0"
      ],
      "expected_awq_modules": 400,
      "extra_sensitivity_protected_modules": 0,
      "fallback": null,
      "foem": null,
      "gc_mode": "on_stage_end",
      "gptaq": null,
      "moe_vram_strategy": "exclusive",
      "moe_vram_strategy_devices": null,
      "mother_revision": "739e3c5b89849f6c238ce1e5b70008612ae42cdd",
      "mse": null,
      "offload_to_disk": true,
      "offload_to_disk_path": null,
      "pack_impl": "cpu",
      "quantizer": [
        "gptqmodel:7.3.2"
      ],
      "recipe": "huihui-qwen38-27b-pure-activation-aware-awq-v2-zero-centered-rmsnorm",
      "scale_search_chunked_activations": false,
      "static_groups": null,
      "structural_modules_to_not_convert": [
        "visual",
        "linear_attn.in_proj_b",
        "linear_attn.in_proj_a",
        "lm_head",
        "mtp"
      ],
      "true_sequential": true,
      "uri": "https://github.com/modelcloud/gptqmodel",
      "wait_for_submodule_finalizers": true
    },
    "method": "awq",
    "pack_dtype": "int32",
    "quant_method": "awq",
    "version": "gemm",
    "zero_point": true,
    "modules_to_not_convert": [
      "visual",
      "linear_attn.in_proj_b",
      "linear_attn.in_proj_a",
      "lm_head",
      "mtp"
    ]
  },
  "awq_module_count": 400,
  "extra_sensitivity_protected_modules": 0,
  "hybrid_module_tree_audit": {
    "status": "PASS",
    "contract": "qkv/z joint AWQ smoothing; retained BF16 a/b in same fanout group",
    "parallel_projection_block": [
      "linear_attn.norm:!",
      "linear_attn.conv1d:!",
      "linear_attn.in_proj_qkv",
      "linear_attn.in_proj_z",
      "linear_attn.in_proj_b:!",
      "linear_attn.in_proj_a:!"
    ],
    "full_block_count": 8
  },
  "zero_centered_rmsnorm_patch": {
    "status": "PASS",
    "class": "transformers.models.qwen3_5.modeling_qwen3_5.Qwen3_5RMSNorm",
    "formula": "norm_weight = (1 + norm_weight) / scale - 1"
  },
  "hybrid_smoothing_compensation": {
    "schema": 2,
    "status": "PASS",
    "method": "joint qkv/z AWQ smoothing plus zero-centered RMSNorm-aware BF16 a/b input-column compensation",
    "zero_centered_rmsnorm": true,
    "formula": "W_ab_compensated = W_ab_source * (1 + norm_source) / (1 + norm_awq)",
    "linear_attention_layers": 48,
    "compensated_tensor_count": 96,
    "touched_shards": [
      "model-00001-of-00003.safetensors"
    ],
    "smoothing_scale": {
      "min": 0.24673694372177124,
      "median": 0.45125672221183777,
      "p99": 0.978907585144043,
      "max": 4.0
    },
    "effective_relative_l2": {
      "min": 0.0016395641723647714,
      "median": 0.0016574671026319265,
      "p95": 0.0016669777687638998,
      "max": 0.0016791383968666196
    }
  },
  "recipe": {
    "status": "PASS",
    "mode": "full",
    "created_utc": "2026-08-25T20:42:14.732344+00:00",
    "stop_after_layer": null,
    "elapsed_seconds_before_save": 6053.315159064,
    "mother": {
      "path": "SOURCE-MANIFEST-BF16-MOTHER.json",
      "sha256": "fed14fbd38a5b6369cf5dc201c6bba3bf4b70ea3887100e0737be4a57fc4a68a",
      "repo": "huihui-ai/Huihui-Qwen3.8-27B-abliterated",
      "resolved_sha": "739e3c5b89849f6c238ce1e5b70008612ae42cdd",
      "file_count": 32,
      "indexed_tensor_count": 1199,
      "indexed_total_size": 55562855904.0
    },
    "host": {
      "name": "NVIDIA GeForce RTX 3090",
      "total_memory": 25296044032,
      "capability": [
        8,
        6
      ],
      "gil_enabled": false
    },
    "calibration_records": 320,
    "calibration_corpus_sha256": "de8c3c856c55afd065ade04929659cf990cde7b8cf5fd846c6ee1867b3a13690",
    "awq_module_count": 400,
    "expected_awq_module_count": 400,
    "extra_sensitivity_protected_modules": 0,
    "hybrid_module_tree_audit": {
      "status": "PASS",
      "contract": "qkv/z joint AWQ smoothing; retained BF16 a/b in same fanout group",
      "parallel_projection_block": [
        "linear_attn.norm:!",
        "linear_attn.conv1d:!",
        "linear_attn.in_proj_qkv",
        "linear_attn.in_proj_z",
        "linear_attn.in_proj_b:!",
        "linear_attn.in_proj_a:!"
      ],
      "full_block_count": 8
    },
    "zero_centered_rmsnorm_patch": {
      "status": "PASS",
      "class": "transformers.models.qwen3_5.modeling_qwen3_5.Qwen3_5RMSNorm",
      "formula": "norm_weight = (1 + norm_weight) / scale - 1"
    },
    "structural_modules_to_not_convert": [
      "visual",
      "linear_attn.in_proj_b",
      "linear_attn.in_proj_a",
      "lm_head",
      "mtp"
    ],
    "sample_count": {
      "min": 202750,
      "max": 202750,
      "average": 202750.0
    },
    "numeric_loss_rows": 336,
    "numeric_loss": {
      "min": 3.862e-07,
      "max": 0.0271780144,
      "average": 0.0011930019714285714
    },
    "activation_audit_sha256": "0f2eef8f907cfd23703af25fc392dd9394d982f4cce39b8fef6e19068083414e",
    "tool": {
      "name": "GPTQModel",
      "version": "7.3.2",
      "commit": "f6f1ab12fe8115f84c52253871ca039ac6c425ee",
      "base_py_sha256": "5550ed2e108e3513a5e7a39748c2d2d42c2b9d78ef3ebea0e7c37657ae8747bd"
    },
    "quantization_script_sha256": "f8a469ccbf47e5689b8fabcd38d1aafb4bd0de594d2b4ba6e4625f4f07c7ae29",
    "hybrid_smoothing_script_sha256": "4e540a56695d225fd8dc164f6b03af59190ef20d403d153cdf2338077790e035",
    "max_rss_kib": 27090320,
    "save_seconds": 84.26336018700022,
    "hybrid_smoothing_compensation": {
      "schema": 2,
      "status": "PASS",
      "method": "joint qkv/z AWQ smoothing plus zero-centered RMSNorm-aware BF16 a/b input-column compensation",
      "zero_centered_rmsnorm": true,
      "formula": "W_ab_compensated = W_ab_source * (1 + norm_source) / (1 + norm_awq)",
      "linear_attention_layers": 48,
      "compensated_tensor_count": 96,
      "touched_shards": [
        "model-00001-of-00003.safetensors"
      ],
      "smoothing_scale": {
        "min": 0.24673694372177124,
        "median": 0.45125672221183777,
        "p99": 0.978907585144043,
        "max": 4.0
      },
      "effective_relative_l2": {
        "min": 0.0016395641723647714,
        "median": 0.0016574671026319265,
        "p95": 0.0016669777687638998,
        "max": 0.0016791383968666196
      }
    }
  }
}