{
  "op": "com.microsoft.MatMulNBitsMlp",
  "tunableSpace": { "TILE_N": [4, 8, 16], "LANES": [8, 16, 32] },
  "cases": [
    {
      "name": "mlp-q4-decode-k2048-n5632",
      "preset": "smoke",
      "vars": { "dtype": "float32" },
      "attrs": { "K": 2048, "N": 5632, "bits": 4, "block_size": 32, "activation": "silu" },
      "inputs": {
        "aT": { "shape": [1, 2048], "dtype": "float32", "dist": "normal", "seed": 8101, "scale": 1 },
        "normScaleT": { "shape": [2048], "dtype": "float32", "dist": "normal", "seed": 8102, "scale": 1 },
        "gateBT": { "shape": [5632, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 8103, "scale": 255 },
        "gateScalesT": { "shape": [5632, 64], "dtype": "float32", "dist": "normal", "seed": 8104, "scale": 0.05 },
        "upBT": { "shape": [5632, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 8105, "scale": 255 },
        "upScalesT": { "shape": [5632, 64], "dtype": "float32", "dist": "normal", "seed": 8106, "scale": 0.05 }
      },
      "outputs": { "yT": { "shape": [1, 5632], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "2 * 5632 * 64 * 16 * 4" }] }
    },
    {
      "name": "mlp-q4-prefill-m64-k2048-n5632",
      "preset": "model",
      "vars": { "dtype": "float32" },
      "attrs": { "K": 2048, "N": 5632, "bits": 4, "block_size": 32, "activation": "silu" },
      "inputs": {
        "aT": { "shape": [64, 2048], "dtype": "float32", "dist": "normal", "seed": 8107, "scale": 1 },
        "normScaleT": { "shape": [2048], "dtype": "float32", "dist": "normal", "seed": 8108, "scale": 1 },
        "gateBT": { "shape": [5632, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 8109, "scale": 255 },
        "gateScalesT": { "shape": [5632, 64], "dtype": "float32", "dist": "normal", "seed": 8110, "scale": 0.05 },
        "upBT": { "shape": [5632, 64, 16], "dtype": "uint8", "dist": "uniform", "seed": 8111, "scale": 255 },
        "upScalesT": { "shape": [5632, 64], "dtype": "float32", "dist": "normal", "seed": 8112, "scale": 0.05 }
      },
      "outputs": { "yT": { "shape": [64, 5632], "dtype": "float32" } },
      "bench": { "metrics": [{ "type": "gflops", "value": "2 * 2 * 64 * 2048 * 5632" }] }
    }
  ]
}