benchmark/a40-linear-kernel-microbench.json

{
  "model": "WaveCut/Krea-2-Turbo-OrbitQuant-W4A4",
  "revision": "49e89c37f51cb394b4ee199196d9df8bcd042475",
  "gpu": "NVIDIA A40",
  "capability": [
    8,
    6
  ],
  "torch": "2.8.0+cu128",
  "layer": "transformer_blocks.0.ff.gate",
  "shape": {
    "rows": 16418,
    "in_features": 6144,
    "out_features": 16384
  },
  "int_mm_chunked": {
    "samples_seconds": [
      0.021558222826570272,
      0.020911033730953932
    ],
    "median_seconds": 0.021234628278762102,
    "peak_allocated_mib": 7793.1279296875,
    "activation_backend": "triton_cuda_packed_w4",
    "native_w4a4_enabled": true,
    "output_mean": 0.001386194839142263,
    "output_std": 1.7071678638458252
  },
  "native_direct_packed": {
    "samples_seconds": [
      0.057071581948548555,
      0.05693301185965538
    ],
    "median_seconds": 0.05700229690410197,
    "peak_allocated_mib": 7459.9912109375,
    "activation_backend": "triton_cuda_packed_w4",
    "native_w4a4_enabled": true,
    "output_mean": 0.0013861971674486995,
    "output_std": 1.7071678638458252
  },
  "direct_over_int_mm": 2.684402861015139
}