{
  "op": "ai.onnx.CausalConvWithState",
  "tunableSpace": { "workgroupSize": [64, 128, 256] },
  "cases": [
    {
      "name": "decode-f16-b1c1024t1k4",
      "preset": "smoke",
      "attrs": { "activation": "silu" },
      "inputs": {
        "inputT": { "shape": [1, 1024, 1], "dtype": "float16", "dist": "normal", "seed": 2751, "scale": 0.2 },
        "weightT": { "shape": [1024, 1, 4], "dtype": "float16", "dist": "normal", "seed": 2752, "scale": 0.1 },
        "biasT": { "shape": [1024], "dtype": "float16", "dist": "normal", "seed": 2753, "scale": 0.05 },
        "pastStateT": { "shape": [1, 1024, 3], "dtype": "float16", "dist": "normal", "seed": 2754, "scale": 0.2 }
      },
      "outputs": {
        "outputT": { "shape": [1, 1024, 1], "dtype": "float16" },
        "presentStateT": { "shape": [1, 1024, 3], "dtype": "float16" }
      },
      "bench": {
        "primary": true,
        "metrics": [
          { "type": "gflops", "value": "2 * numel(shapes.outputT) * dim(shapes.weightT, 2)" },
          {
            "type": "bandwidth",
            "value": "(numel(shapes.inputT) + numel(shapes.weightT) + numel(shapes.biasT) + numel(shapes.pastStateT) + numel(shapes.outputT) + numel(shapes.presentStateT)) * 2"
          }
        ]
      }
    },
    {
      "name": "prefill-f32-b2c32t256k4",
      "preset": "smoke",
      "attrs": { "activation": "none" },
      "inputs": {
        "inputT": { "shape": [2, 32, 256], "dtype": "float32", "dist": "normal", "seed": 2755, "scale": 0.2 },
        "weightT": { "shape": [32, 1, 4], "dtype": "float32", "dist": "normal", "seed": 2756, "scale": 0.1 }
      },
      "outputs": {
        "outputT": { "shape": [2, 32, 256], "dtype": "float32" },
        "presentStateT": { "shape": [2, 32, 3], "dtype": "float32" }
      },
      "bench": {
        "primary": true,
        "metrics": [
          { "type": "gflops", "value": "2 * numel(shapes.outputT) * dim(shapes.weightT, 2)" },
          {
            "type": "bandwidth",
            "value": "(numel(shapes.inputT) + numel(shapes.weightT) + numel(shapes.outputT) + numel(shapes.presentStateT)) * 4"
          }
        ]
      }
    },
    {
      "name": "large-kernel-f32-b2c1024t512k128",
      "preset": "stress",
      "tunableSpace": { "tiledWorkgroupSize": [64, 128, 256] },
      "attrs": { "activation": "none" },
      "inputs": {
        "inputT": { "shape": [2, 1024, 512], "dtype": "float32", "dist": "normal", "seed": 2757, "scale": 0.2 },
        "weightT": { "shape": [1024, 1, 128], "dtype": "float32", "dist": "normal", "seed": 2758, "scale": 0.05 }
      },
      "outputs": {
        "outputT": { "shape": [2, 1024, 512], "dtype": "float32", "dist": "empty" },
        "presentStateT": { "shape": [2, 1024, 127], "dtype": "float32", "dist": "empty" }
      },
      "bench": {
        "primary": true,
        "metrics": [
          { "type": "gflops", "value": "2 * numel(shapes.outputT) * dim(shapes.weightT, 2)" },
          {
            "type": "bandwidth",
            "value": "(numel(shapes.inputT) + numel(shapes.weightT) + numel(shapes.outputT) + numel(shapes.presentStateT)) * 4"
          }
        ]
      }
    }
  ]
}