{
  "domain": "com.microsoft",
  "name": "LinearAttentionGate",
  "sinceVersion": 1,
  "inputs": {
    "aT": { "onnx": "a", "dtype": "T" },
    "dtBiasT": { "onnx": "dt_bias", "dtype": "TF", "rank": 1 },
    "decayScaleT": { "onnx": "decay_scale", "dtype": "TF", "rank": 1 },
    "bT": { "onnx": "b", "dtype": "T", "optional": true }
  },
  "outputs": {
    "decayT": { "onnx": "decay", "dtype": "T", "rank": "ranks.aT", "shape": "shapes.aT" },
    "betaT": { "onnx": "beta", "dtype": "T", "rank": "ranks.aT", "optional": true, "shape": "shapes.aT" }
  },
  "typeConstraints": { "T": ["float32", "float16"], "TF": ["float32"] },
  "tunables": { "WORKGROUP_SIZE": { "default": 64 } },
  "derive": {
    "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
    "foldedDispatchCapacity": "min(device.limits.maxComputeWorkgroupsPerDimension, 65535) * min(device.limits.maxComputeWorkgroupsPerDimension, 65535)",
    "numHeads": "dim(shapes.aT, ranks.aT - 1)",
    "gateCount": "numel(shapes.aT)",
    "headsVec4": "numHeads / 4",
    "gateVec4Count": "gateCount / 4",
    "gateDtype": "tensorDtypes.aT",
    "gateDtypeOk": "(gateDtype == \"float32\" or gateDtype == \"float16\") and f16Ok(dtypes.T)",
    "paramsOk": "ranks.aT >= 1 and numHeads > 0 and ranks.dtBiasT == 1 and ranks.decayScaleT == 1 and tensorDtypes.dtBiasT == \"float32\" and tensorDtypes.decayScaleT == \"float32\" and dim(shapes.dtBiasT, 0) == numHeads and dim(shapes.decayScaleT, 0) == numHeads",
    "tensorContract": "gateDtypeOk and paramsOk and sameShape(shapes.decayT, shapes.aT) and tensorDtypes.decayT == gateDtype",
    "betaContract": "tensorContract and present.bT and present.betaT and sameShape(shapes.bT, shapes.aT) and sameShape(shapes.betaT, shapes.aT) and tensorDtypes.bT == gateDtype and tensorDtypes.betaT == gateDtype",
    "decayOnlyContract": "tensorContract and not present.betaT",
    "workgroupFits": "tunables.WORKGROUP_SIZE > 0 and tunables.WORKGROUP_SIZE <= deviceWorkgroupCap",
    "scalarDispatchFits": "ceilDiv(gateCount, tunables.WORKGROUP_SIZE) <= foldedDispatchCapacity",
    "vec4DispatchFits": "numHeads % 4 == 0 and ceilDiv(gateVec4Count, tunables.WORKGROUP_SIZE) <= foldedDispatchCapacity",
    "workgroupSize": "tunables.WORKGROUP_SIZE"
  },
  "when": ["workgroupFits"],
  "bindings": {
    "a": { "arg": "aT", "elementType": "$gateElement", "length": "$gateItems" },
    "dt_bias": { "arg": "dtBiasT", "elementType": "$paramElement", "length": "$headItems" },
    "decay_scale": { "arg": "decayScaleT", "elementType": "$paramElement", "length": "$headItems" },
    "b": { "arg": "bT", "elementType": "$gateElement", "length": "$gateItems" },
    "decay": { "arg": "decayT", "elementType": "$gateElement", "length": "$gateItems" },
    "beta": { "arg": "betaT", "elementType": "$gateElement", "length": "$gateItems" }
  },
  "variants": [
    {
      "id": "vec4_gate_beta",
      "priority": 30,
      "when": ["betaContract", "vec4DispatchFits"],
      "derive": {
        "vectorized": true,
        "hasBeta": "present.betaT",
        "gateElement": "\"vec4<f16>\" if gateDtype == \"float16\" else \"vec4<f32>\"",
        "paramElement": "\"vec4<f32>\"",
        "headItems": "headsVec4",
        "gateItems": "gateVec4Count"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.Vec4GateBeta",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": ["a", "dt_bias", "decay_scale", "b", "decay", "beta"],
          "dispatch": {
            "x": "min(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "y": "ceilDiv(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "vec4_gate",
      "priority": 20,
      "when": ["decayOnlyContract", "vec4DispatchFits"],
      "derive": {
        "vectorized": true,
        "hasBeta": "present.betaT",
        "gateElement": "\"vec4<f16>\" if gateDtype == \"float16\" else \"vec4<f32>\"",
        "paramElement": "\"vec4<f32>\"",
        "headItems": "headsVec4",
        "gateItems": "gateVec4Count"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.Vec4Gate",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": ["a", "dt_bias", "decay_scale", "decay"],
          "dispatch": {
            "x": "min(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "y": "ceilDiv(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "scalar_gate_beta",
      "priority": 10,
      "when": ["betaContract", "scalarDispatchFits"],
      "derive": {
        "vectorized": false,
        "hasBeta": "present.betaT",
        "gateElement": "\"f16\" if gateDtype == \"float16\" else \"f32\"",
        "paramElement": "\"f32\"",
        "headItems": "numHeads",
        "gateItems": "gateCount"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.ScalarGateBeta",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": ["a", "dt_bias", "decay_scale", "b", "decay", "beta"],
          "dispatch": {
            "x": "min(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "y": "ceilDiv(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "z": 1
          }
        }
      ]
    },
    {
      "id": "scalar_gate",
      "priority": 0,
      "when": ["decayOnlyContract", "scalarDispatchFits"],
      "derive": {
        "vectorized": false,
        "hasBeta": "present.betaT",
        "gateElement": "\"f16\" if gateDtype == \"float16\" else \"f32\"",
        "paramElement": "\"f32\"",
        "headItems": "numHeads",
        "gateItems": "gateCount"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.ScalarGate",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": ["a", "dt_bias", "decay_scale", "decay"],
          "dispatch": {
            "x": "min(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "y": "ceilDiv(ceilDiv((gateItems), (tunables.WORKGROUP_SIZE)), 65535)",
            "z": 1
          }
        }
      ]
    }
  ]
}