outputs/kernel_bench_quick.json

{
  "device": "cpu",
  "dtype": "torch.float32",
  "rows": [
    {
      "L": 64,
      "D": 1,
      "mha_ms": 0.10488299885764718,
      "mla_ms": 20.717017899733037
    },
    {
      "L": 64,
      "D": 4,
      "mha_ms": 0.23264494957402349,
      "mla_ms": 20.071504899533466
    },
    {
      "L": 64,
      "D": 16,
      "mha_ms": 0.8265429467428476,
      "mla_ms": 23.22479714639485
    },
    {
      "L": 256,
      "D": 1,
      "mha_ms": 1.6084900998976082,
      "mla_ms": 50.07811555406079
    },
    {
      "L": 256,
      "D": 4,
      "mha_ms": 1.5490520978346467,
      "mla_ms": 46.594679151894525
    },
    {
      "L": 256,
      "D": 16,
      "mha_ms": 1.7690911539830267,
      "mla_ms": 54.35818649712019
    }
  ]
}