{
  "comparison": {
    "elements": 12288,
    "lhs": "fused_turboquant_gemv",
    "max_abs": 0.0078125,
    "max_abs_index": 1290,
    "max_rel": 29.0944881439209,
    "max_rel_index": 1173,
    "mean_abs": 0.00054884486598894,
    "mean_rel": 0.014326672850719055,
    "rhs": "bulk_dequant_cublas",
    "rhs_rms": 0.414406758531513,
    "rmse": 0.0010549820192550035,
    "rmse_over_rhs_rms": 0.002545764511644129
  },
  "first8": [
    {
      "abs_err": 0.0009765625,
      "bulk_dequant_cublas": -0.2060546875,
      "fused_gemv": -0.20703125,
      "index": 0,
      "rel_err": 0.004739336669445038
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.267578125,
      "fused_gemv": -0.267578125,
      "index": 1,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.10888671875,
      "fused_gemv": -0.10888671875,
      "index": 2,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.40625,
      "fused_gemv": 0.40625,
      "index": 3,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.220703125,
      "fused_gemv": -0.220703125,
      "index": 4,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.400390625,
      "fused_gemv": -0.400390625,
      "index": 5,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.038330078125,
      "fused_gemv": 0.038330078125,
      "index": 6,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0009765625,
      "bulk_dequant_cublas": -0.138671875,
      "fused_gemv": -0.1396484375,
      "index": 7,
      "rel_err": 0.007042253389954567
    }
  ],
  "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": 12288
  },
  "seed": 1592594996,
  "tensor_base": "model.language_model.layers.10.mlp.gate_proj"
}
