{
  "op": "com.microsoft.MRotaryEmbedding",
  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
  "cases": [
    {
      "name": "mrope-f32-smoke-b1h4s64d64",
      "preset": "smoke",
      "attrs": {
        "mrope_layout": 0,
        "interleaved": 0,
        "mrope_section": [16, 8, 8],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 16384, "rotPairs": 8192 },
      "inputs": {
        "x": { "shape": [1, 4, 64, 64], "dtype": "float32", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": { "shape": [3, 1, 64], "dtype": "uint32", "dist": "uniform", "seed": 1001, "min": 0, "max": 127 },
        "cos": { "shape": [128, 32], "dtype": "float32", "dist": "uniform", "seed": 1002, "scale": 0.1, "offset": 0.9 },
        "sin": { "shape": [128, 32], "dtype": "float32", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [1, 4, 64, 64], "dtype": "float32" } },
      "bench": {
        "primary": false,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 4 * 2 + args.rotPairs * 4 * 2" }]
      }
    },
    {
      "name": "qwen3-vl-2b-decode-q-h16-d128",
      "preset": "model",
      "attrs": {
        "mrope_layout": 1,
        "interleaved": 0,
        "mrope_section": [24, 20, 20],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 2048, "rotPairs": 1024 },
      "provenance": {
        "model": "Qwen/Qwen3-VL-2B-Instruct",
        "notes": "Qwen3-VL interleaved M-RoPE over the query projection: 16 heads x head_dim 128, mrope_section [24, 20, 20] summing to head_dim/2. Cache rows cover the model's position budget."
      },
      "inputs": {
        "x": { "shape": [1, 16, 1, 128], "dtype": "float32", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": {
          "shape": [3, 1, 1],
          "dtype": "uint32",
          "dist": "uniform",
          "seed": 1001,
          "min": 0,
          "max": 32767
        },
        "cos": {
          "shape": [32768, 64],
          "dtype": "float32",
          "dist": "uniform",
          "seed": 1002,
          "scale": 0.1,
          "offset": 0.9
        },
        "sin": { "shape": [32768, 64], "dtype": "float32", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [1, 16, 1, 128], "dtype": "float32" } },
      "bench": {
        "primary": true,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 4 * 2 + args.rotPairs * 4 * 2" }]
      }
    },
    {
      "name": "qwen3-vl-2b-prefill-t512-q-h16-d128",
      "preset": "model",
      "attrs": {
        "mrope_layout": 1,
        "interleaved": 0,
        "mrope_section": [24, 20, 20],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 1048576, "rotPairs": 524288 },
      "provenance": {
        "model": "Qwen/Qwen3-VL-2B-Instruct",
        "notes": "Qwen3-VL interleaved M-RoPE over the query projection: 16 heads x head_dim 128, mrope_section [24, 20, 20] summing to head_dim/2. Cache rows cover the model's position budget."
      },
      "inputs": {
        "x": { "shape": [1, 16, 512, 128], "dtype": "float32", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": {
          "shape": [3, 1, 512],
          "dtype": "uint32",
          "dist": "uniform",
          "seed": 1001,
          "min": 0,
          "max": 32767
        },
        "cos": {
          "shape": [32768, 64],
          "dtype": "float32",
          "dist": "uniform",
          "seed": 1002,
          "scale": 0.1,
          "offset": 0.9
        },
        "sin": { "shape": [32768, 64], "dtype": "float32", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [1, 16, 512, 128], "dtype": "float32" } },
      "bench": {
        "primary": false,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 4 * 2 + args.rotPairs * 4 * 2" }]
      }
    },
    {
      "name": "qwen3-vl-2b-decode-k-h8-d128",
      "preset": "model",
      "attrs": {
        "mrope_layout": 1,
        "interleaved": 0,
        "mrope_section": [24, 20, 20],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 1024, "rotPairs": 512 },
      "provenance": {
        "model": "Qwen/Qwen3-VL-2B-Instruct",
        "notes": "Same layer's key projection: 8 kv heads, so half the query work and the smallest dispatch M-RoPE sees in decode."
      },
      "inputs": {
        "x": { "shape": [1, 8, 1, 128], "dtype": "float32", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": {
          "shape": [3, 1, 1],
          "dtype": "uint32",
          "dist": "uniform",
          "seed": 1001,
          "min": 0,
          "max": 32767
        },
        "cos": {
          "shape": [32768, 64],
          "dtype": "float32",
          "dist": "uniform",
          "seed": 1002,
          "scale": 0.1,
          "offset": 0.9
        },
        "sin": { "shape": [32768, 64], "dtype": "float32", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [1, 8, 1, 128], "dtype": "float32" } },
      "bench": {
        "primary": false,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 4 * 2 + args.rotPairs * 4 * 2" }]
      }
    },
    {
      "name": "qwen2-5-vl-3b-prefill-t512-q-h16-d128",
      "preset": "model",
      "attrs": {
        "mrope_layout": 0,
        "interleaved": 0,
        "mrope_section": [16, 24, 24],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 1048576, "rotPairs": 524288 },
      "provenance": {
        "model": "Qwen/Qwen2.5-VL-3B-Instruct",
        "notes": "Qwen2.5-VL sectioned M-RoPE, mrope_section [16, 24, 24]. The sectioned branch reads three contiguous chunks where the interleaved branch strides by 3, so the two layouts have different cache access patterns at the same shape."
      },
      "inputs": {
        "x": { "shape": [1, 16, 512, 128], "dtype": "float32", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": {
          "shape": [3, 1, 512],
          "dtype": "uint32",
          "dist": "uniform",
          "seed": 1001,
          "min": 0,
          "max": 32767
        },
        "cos": {
          "shape": [32768, 64],
          "dtype": "float32",
          "dist": "uniform",
          "seed": 1002,
          "scale": 0.1,
          "offset": 0.9
        },
        "sin": { "shape": [32768, 64], "dtype": "float32", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [1, 16, 512, 128], "dtype": "float32" } },
      "bench": {
        "primary": false,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 4 * 2 + args.rotPairs * 4 * 2" }]
      }
    },
    {
      "name": "qwen3-vl-2b-prefill-t512-q-h16-d128-f16",
      "preset": "model",
      "attrs": {
        "mrope_layout": 1,
        "interleaved": 0,
        "mrope_section": [24, 20, 20],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 1048576, "rotPairs": 524288 },
      "provenance": {
        "model": "Qwen/Qwen3-VL-2B-Instruct",
        "notes": "f16 activations at the same Qwen3-VL prefill shape."
      },
      "inputs": {
        "x": { "shape": [1, 16, 512, 128], "dtype": "float16", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": {
          "shape": [3, 1, 512],
          "dtype": "uint32",
          "dist": "uniform",
          "seed": 1001,
          "min": 0,
          "max": 32767
        },
        "cos": {
          "shape": [32768, 64],
          "dtype": "float16",
          "dist": "uniform",
          "seed": 1002,
          "scale": 0.1,
          "offset": 0.9
        },
        "sin": { "shape": [32768, 64], "dtype": "float16", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [1, 16, 512, 128], "dtype": "float16" } },
      "bench": {
        "primary": false,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 2 * 2 + args.rotPairs * 2 * 2" }]
      }
    },
    {
      "name": "mrope-f32-stress-b4h32s2048d128",
      "preset": "stress",
      "attrs": {
        "mrope_layout": 1,
        "interleaved": 0,
        "mrope_section": [24, 20, 20],
        "rotary_embedding_dim": 0,
        "scale": 1
      },
      "vars": { "elems": 33554432, "rotPairs": 16777216 },
      "inputs": {
        "x": { "shape": [4, 32, 2048, 128], "dtype": "float32", "dist": "normal", "seed": 1000, "scale": 0.2 },
        "positionIds": {
          "shape": [3, 4, 2048],
          "dtype": "uint32",
          "dist": "uniform",
          "seed": 1001,
          "min": 0,
          "max": 32767
        },
        "cos": {
          "shape": [32768, 64],
          "dtype": "float32",
          "dist": "uniform",
          "seed": 1002,
          "scale": 0.1,
          "offset": 0.9
        },
        "sin": { "shape": [32768, 64], "dtype": "float32", "dist": "uniform", "seed": 1003, "scale": 0.1 }
      },
      "outputs": { "y": { "shape": [4, 32, 2048, 128], "dtype": "float32" } },
      "bench": {
        "primary": false,
        "metrics": [{ "type": "bandwidth", "value": "args.elems * 4 * 2 + args.rotPairs * 4 * 2" }]
      }
    }
  ]
}