{
  "comparison": {
    "elements": 1024,
    "lhs": "fused_turboquant_gemv",
    "max_abs": 0.015625,
    "max_abs_index": 282,
    "max_rel": 0.3918918967247009,
    "max_rel_index": 166,
    "mean_abs": 0.0007867664098739624,
    "mean_rel": 0.006071774091196858,
    "rhs": "bulk_dequant_cublas",
    "rhs_rms": 0.5948146975992104,
    "rmse": 0.001648442469949082,
    "rmse_over_rhs_rms": 0.0027713546363304768
  },
  "first8": [
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.8203125,
      "fused_gemv": 0.8203125,
      "index": 0,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.474609375,
      "fused_gemv": 0.474609375,
      "index": 1,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.5390625,
      "fused_gemv": -0.5390625,
      "index": 2,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.90234375,
      "fused_gemv": -0.90234375,
      "index": 3,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.40625,
      "fused_gemv": -0.40625,
      "index": 4,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": 0.734375,
      "fused_gemv": 0.734375,
      "index": 5,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.0,
      "bulk_dequant_cublas": -0.55859375,
      "fused_gemv": -0.55859375,
      "index": 6,
      "rel_err": 0.0
    },
    {
      "abs_err": 0.001953125,
      "bulk_dequant_cublas": -0.3203125,
      "fused_gemv": -0.322265625,
      "index": 7,
      "rel_err": 0.006097560748457909
    }
  ],
  "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": 1024
  },
  "seed": 1592594996,
  "tensor_base": "model.language_model.layers.3.self_attn.k_proj"
}
