{
  "comparison": {
    "elements": 4096,
    "lhs": "fused_turboquant_gemv",
    "max_abs": 0.015625,
    "max_abs_index": 748,
    "max_rel": 2.744680881500244,
    "max_rel_index": 2575,
    "mean_abs": 0.000923495004826691,
    "mean_rel": 0.010134102004030865,
    "rhs": "bulk_dequant_cublas",
    "rhs_rms": 0.7002984414198531,
    "rmse": 0.0017891602000170233,
    "rmse_over_rhs_rms": 0.002554853893990548
  },
  "first8": [
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.287109375,
      "fused_gemv": 0.287109375,
      "index": 0,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.578125,
      "fused_gemv": -0.578125,
      "index": 1,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.50390625,
      "fused_gemv": -0.50390625,
      "index": 2,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.001953125,
      "bulk_dequant_cublas": 0.240234375,
      "fused_gemv": 0.2421875,
      "index": 3,
      "rel_err": 0.008130080997943878
    },
    {
      "abs_err": 0.00048828125,
      "bulk_dequant_cublas": -0.05712890625,
      "fused_gemv": -0.0576171875,
      "index": 4,
      "rel_err": 0.008547008968889713
    },
    {
      "abs_err": 0.001953125,
      "bulk_dequant_cublas": 0.34375,
      "fused_gemv": 0.345703125,
      "index": 5,
      "rel_err": 0.005681818351149559
    },
    {
      "abs_err": 0.00244140625,
      "bulk_dequant_cublas": 0.06787109375,
      "fused_gemv": 0.0703125,
      "index": 6,
      "rel_err": 0.035971224308013916
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.1533203125,
      "fused_gemv": -0.1533203125,
      "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,
      12288
    ]
  },
  "model_path": "/home/ckl/.cache/modelscope/hub/Qwen/Qwen3___5-9B-TQ4",
  "projection": {
    "bits": 4,
    "cols": 12288,
    "group_size": 128,
    "num_groups": 96,
    "packed_cols": 6144,
    "rows": 4096
  },
  "seed": 1592594996,
  "tensor_base": "model.language_model.layers.3.mlp.down_proj"
}
