{
  "cases": [
    {
      "name": "row_slices",
      "vars": { "indexCount": 128, "cols": 768, "dtype": "float32" },
      "inputs": {
        "data": { "dtype": "float32", "shape": [4096, 768] },
        "indices": { "dtype": "int32", "shape": [128, 1], "min": 0, "max": 4096 }
      },
      "outputs": { "output": { "dtype": "float32", "shape": [128, 768] } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "embedding-row-slices-f16-32000x512-idx1024",
      "preset": "smoke",
      "vars": { "rows": 32000, "cols": 512, "indexCount": 1024, "dtype": "float16" },
      "inputs": {
        "data": { "dtype": "float16", "shape": [32000, 512], "dist": "normal", "seed": 125, "scale": 0.2 },
        "indices": { "dtype": "int32", "shape": [1024, 1], "dist": "linearMod", "seed": 125, "step": 29, "mod": 32000 }
      },
      "outputs": { "output": { "dtype": "float16", "shape": [1024, 512] } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "alignment-cliff-f16-32000x513-idx1024-unaligned-scalar",
      "preset": "smoke",
      "vars": { "rows": 32000, "cols": 513, "indexCount": 1024, "dtype": "float16" },
      "inputs": {
        "data": { "dtype": "float16", "shape": [32000, 513], "dist": "normal", "seed": 125, "scale": 0.2 },
        "indices": { "dtype": "int32", "shape": [1024, 1], "dist": "linearMod", "seed": 125, "step": 29, "mod": 32000 }
      },
      "outputs": { "output": { "dtype": "float16", "shape": [1024, 513] } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "missing-regime-batch1-wide-row-scalar-64x4096x768",
      "preset": "stress",
      "provenance": { "notes": "Stress-only capacity case: declared tensors occupy 792 MiB of GPU storage." },
      "vars": { "batch": 64, "rows": 4096, "cols": 768, "indexCount": 128, "dtype": "float32" },
      "attrs": { "batch_dims": 1 },
      "inputs": {
        "data": { "dtype": "float32", "shape": [64, 4096, 768], "dist": "normal", "seed": 126, "scale": 0.2 },
        "indices": {
          "dtype": "int32",
          "shape": [64, 128, 1],
          "dist": "linearMod",
          "seed": 126,
          "step": 31,
          "mod": 4096
        }
      },
      "outputs": { "output": { "dtype": "float32", "shape": [64, 128, 768] } },
      "bench": {
        "metrics": [
          { "type": "bandwidth", "value": "args.batch * args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }
        ]
      }
    },
    {
      "name": "narrow-row-low-occupancy-vec4-100000x8-idx65536",
      "preset": "smoke",
      "vars": { "rows": 100000, "cols": 8, "indexCount": 65536, "dtype": "float32" },
      "inputs": {
        "data": { "dtype": "float32", "shape": [100000, 8], "dist": "normal", "seed": 127, "scale": 0.2 },
        "indices": {
          "dtype": "int32",
          "shape": [65536, 1],
          "dist": "linearMod",
          "seed": 127,
          "step": 37,
          "mod": 100000
        }
      },
      "outputs": { "output": { "dtype": "float32", "shape": [65536, 8] } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "dispatch-cliff-vec4-4096x768-idx131072-gridstride",
      "preset": "stress",
      "provenance": { "notes": "Stress-only dispatch/capacity case: declared tensors occupy 396 MiB of GPU storage." },
      "vars": { "rows": 4096, "cols": 768, "indexCount": 131072, "dtype": "float32" },
      "inputs": {
        "data": { "dtype": "float32", "shape": [4096, 768], "dist": "normal", "seed": 128, "scale": 0.2 },
        "indices": { "dtype": "int32", "shape": [131072, 1], "dist": "linearMod", "seed": 128, "step": 41, "mod": 4096 }
      },
      "outputs": { "output": { "dtype": "float32", "shape": [131072, 768] } },
      "bench": {
        "metrics": [{ "type": "bandwidth", "value": "args.indexCount * args.cols * dtypeBytes(args.dtype) * 2" }]
      }
    },
    {
      "name": "gathernd-f32-batch3-b2x4x8-rows512-width512-kv-reorder-pathology",
      "preset": "stress",
      "provenance": {
        "kind": "synthetic",
        "description": "Realistic rank-5 KV-cache reorder with three batch dimensions; an equal-element row_fast_vec4 case isolates the batch3 route overhead."
      },
      "vars": { "dtype": "float32", "count": 16777216 },
      "attrs": { "batch_dims": 3 },
      "inputs": {
        "data": { "dtype": "float32", "shape": [2, 4, 8, 512, 512], "dist": "normal", "seed": 941, "scale": 1 },
        "indices": {
          "dtype": "int32",
          "shape": [2, 4, 8, 512, 1],
          "dist": "linearMod",
          "seed": 942,
          "step": 17,
          "mod": 512
        }
      },
      "outputs": { "output": { "dtype": "float32", "shape": [2, 4, 8, 512, 512] } },
      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
    }
  ]
}