{
  "comparison": {
    "elements": 4096,
    "lhs": "fused_turboquant_gemv",
    "max_abs": 0.0078125,
    "max_abs_index": 431,
    "max_rel": 50.284671783447266,
    "max_rel_index": 2347,
    "mean_abs": 0.0007508644484914839,
    "mean_rel": 0.021170889410086602,
    "rhs": "bulk_dequant_cublas",
    "rhs_rms": 0.594118056326469,
    "rmse": 0.0014473373169397725,
    "rmse_over_rhs_rms": 0.0024361106374865974
  },
  "first8": [
    {
      "abs_err": 0.001953125,
      "bulk_dequant_cublas": -0.162109375,
      "fused_gemv": -0.1640625,
      "index": 0,
      "rel_err": 0.012048192322254181
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.5234375,
      "fused_gemv": 0.5234375,
      "index": 1,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0029296875,
      "bulk_dequant_cublas": -0.111328125,
      "fused_gemv": -0.1083984375,
      "index": 2,
      "rel_err": 0.02631578966975212
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.7421875,
      "fused_gemv": -0.7421875,
      "index": 3,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0009765625,
      "bulk_dequant_cublas": -0.10205078125,
      "fused_gemv": -0.10107421875,
      "index": 4,
      "rel_err": 0.009569377638399601
    },
    {
      "abs_err": 0.00390625,
      "bulk_dequant_cublas": -0.890625,
      "fused_gemv": -0.89453125,
      "index": 5,
      "rel_err": 0.004385964944958687
    },
    {
      "abs_err": 0.000244140625,
      "bulk_dequant_cublas": 0.00994873046875,
      "fused_gemv": 0.00970458984375,
      "index": 6,
      "rel_err": 0.024539876729249954
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.298828125,
      "fused_gemv": 0.298828125,
      "index": 7,
      "rel_err": 0.0
    }
  ],
  "input": {
    "first8": [
      0.60546875,
      0.466796875,
      -0.35546875,
      -0.48828125,
      -0.984375,
      0.216796875,
      -0.12890625,
      -0.52734375
    ],
    "shape": [
      1,
      4096
    ]
  },
  "model_path": "/home/ckl/.cache/modelscope/hub/Qwen/Qwen3___5-9B-TQ4",
  "projection": {
    "bits": 4,
    "cols": 4096,
    "group_size": 128,
    "num_groups": 32,
    "packed_cols": 2048,
    "rows": 4096
  },
  "seed": 1592594996,
  "tensor_base": "model.language_model.layers.3.self_attn.o_proj"
}
