{
  "domain": "com.microsoft",
  "name": "LinearAttentionGate",
  "sinceVersion": 1,
  "description": "Fuses the gate projections used by `com.microsoft.LinearAttention`'s gated-delta recurrence: `decay = decay_scale * softplus(a + dt_bias)` and, when requested, `beta = sigmoid(b)`. The last input axis is the head axis; `dt_bias` and `decay_scale` are float32 per-head vectors. Gate arithmetic is performed in float32 and narrowed only on store. Requesting `beta` requires `b`; an unconsumed `b` is permitted when `beta` is omitted.",
  "inputs": [
    {
      "role": "a",
      "dtype": "T",
      "description": "Decay gate projection with shape `(B, T, H)`. Any rank of at least 1 is accepted; the last axis is the head count and the leading axes are folded."
    },
    {
      "role": "dt_bias",
      "dtype": "TF",
      "rank": 1,
      "description": "Per-head float32 bias added to `a`, with shape (H)."
    },
    {
      "role": "decay_scale",
      "dtype": "TF",
      "rank": 1,
      "description": "Per-head float32 multiplier applied to `softplus(a + dt_bias)`, with shape `(H)`. For gated DeltaNet this is `-exp(A_log)`."
    },
    {
      "role": "b",
      "dtype": "T",
      "optional": true,
      "description": "Update-rate projection with the same shape as `a` when `beta` is requested. It is accepted but unused when `beta` is omitted."
    }
  ],
  "outputs": [
    {
      "role": "decay",
      "dtype": "T",
      "rank": "ranks.aT",
      "shape": "shapes.aT",
      "description": "`decay_scale * softplus(a + dt_bias)`, with the same shape as `a`."
    },
    {
      "role": "beta",
      "dtype": "T",
      "rank": "ranks.aT",
      "optional": true,
      "shape": "shapes.aT",
      "description": "sigmoid(b), with the same shape as `a`. Requires the `b` input."
    }
  ],
  "typeConstraints": { "T": ["float32", "float16"], "TF": ["float32"] },
  "tunables": { "WORKGROUP_SIZE": 64 },
  "args": {
    "aT": { "kind": "tensor", "semantic": "a", "role": "input" },
    "dtBiasT": { "kind": "tensor", "semantic": "dt_bias", "role": "weights" },
    "decayScaleT": { "kind": "tensor", "semantic": "decay_scale", "role": "weights" },
    "bT": { "kind": "tensor", "semantic": "b", "role": "input", "required": false },
    "decayT": { "kind": "tensor", "semantic": "decay", "role": "output" },
    "betaT": { "kind": "tensor", "semantic": "beta", "role": "output", "required": false }
  },
  "derive": {
    "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
    "foldedDispatchCapacity": "device.limits.maxComputeWorkgroupsPerDimension * device.limits.maxComputeWorkgroupsPerDimension",
    "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"
  },
  "bindingSets": {
    "decayOnlyIo": [
      {
        "name": "a",
        "arg": "aT",
        "semantic": "a",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$gateElement",
        "length": "$gateItems"
      },
      {
        "name": "dt_bias",
        "arg": "dtBiasT",
        "semantic": "dt_bias",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$paramElement",
        "length": "$headItems"
      },
      {
        "name": "decay_scale",
        "arg": "decayScaleT",
        "semantic": "decay_scale",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$paramElement",
        "length": "$headItems"
      },
      {
        "name": "decay",
        "arg": "decayT",
        "semantic": "decay",
        "buffer": { "type": "storage" },
        "elementType": "$gateElement",
        "length": "$gateItems"
      }
    ],
    "withBetaIo": [
      {
        "name": "a",
        "arg": "aT",
        "semantic": "a",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$gateElement",
        "length": "$gateItems"
      },
      {
        "name": "dt_bias",
        "arg": "dtBiasT",
        "semantic": "dt_bias",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$paramElement",
        "length": "$headItems"
      },
      {
        "name": "decay_scale",
        "arg": "decayScaleT",
        "semantic": "decay_scale",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$paramElement",
        "length": "$headItems"
      },
      {
        "name": "b",
        "arg": "bT",
        "semantic": "b",
        "buffer": { "type": "read-only-storage" },
        "elementType": "$gateElement",
        "length": "$gateItems"
      },
      {
        "name": "decay",
        "arg": "decayT",
        "semantic": "decay",
        "buffer": { "type": "storage" },
        "elementType": "$gateElement",
        "length": "$gateItems"
      },
      {
        "name": "beta",
        "arg": "betaT",
        "semantic": "beta",
        "buffer": { "type": "storage" },
        "elementType": "$gateElement",
        "length": "$gateItems"
      }
    ]
  },
  "variants": [
    {
      "id": "vec4_gate_beta",
      "priority": 30,
      "when": ["betaContract", "workgroupFits", "vec4DispatchFits"],
      "constants": {
        "vectorized": true,
        "hasBeta": true,
        "usesF16": "gateDtype == \"float16\"",
        "gateElement": "\"vec4<f16>\" if gateDtype == \"float16\" else \"vec4<f32>\"",
        "paramElement": "\"vec4<f32>\"",
        "headItems": "headsVec4",
        "gateItems": "gateVec4Count",
        "workgroupSize": "tunables.WORKGROUP_SIZE"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.Vec4GateBeta",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": "withBetaIo",
          "dispatch": { "threads": "gateVec4Count", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    },
    {
      "id": "vec4_gate",
      "priority": 20,
      "when": ["decayOnlyContract", "workgroupFits", "vec4DispatchFits"],
      "constants": {
        "vectorized": true,
        "hasBeta": false,
        "usesF16": "gateDtype == \"float16\"",
        "gateElement": "\"vec4<f16>\" if gateDtype == \"float16\" else \"vec4<f32>\"",
        "paramElement": "\"vec4<f32>\"",
        "headItems": "headsVec4",
        "gateItems": "gateVec4Count",
        "workgroupSize": "tunables.WORKGROUP_SIZE"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.Vec4Gate",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": "decayOnlyIo",
          "dispatch": { "threads": "gateVec4Count", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    },
    {
      "id": "scalar_gate_beta",
      "priority": 10,
      "when": ["betaContract", "workgroupFits", "scalarDispatchFits"],
      "constants": {
        "vectorized": false,
        "hasBeta": true,
        "usesF16": "gateDtype == \"float16\"",
        "gateElement": "\"f16\" if gateDtype == \"float16\" else \"f32\"",
        "paramElement": "\"f32\"",
        "headItems": "numHeads",
        "gateItems": "gateCount",
        "workgroupSize": "tunables.WORKGROUP_SIZE"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.ScalarGateBeta",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": "withBetaIo",
          "dispatch": { "threads": "gateCount", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    },
    {
      "id": "scalar_gate",
      "priority": 0,
      "when": ["decayOnlyContract", "workgroupFits", "scalarDispatchFits"],
      "constants": {
        "vectorized": false,
        "hasBeta": false,
        "usesF16": "gateDtype == \"float16\"",
        "gateElement": "\"f16\" if gateDtype == \"float16\" else \"f32\"",
        "paramElement": "\"f32\"",
        "headItems": "numHeads",
        "gateItems": "gateCount",
        "workgroupSize": "tunables.WORKGROUP_SIZE"
      },
      "passes": [
        {
          "id": "main",
          "name": "LinearAttentionGate.ScalarGate",
          "shader": "linear-attention-gate.wgsl.jinja",
          "bindings": "decayOnlyIo",
          "dispatch": { "threads": "gateCount", "workgroupSize": "tunables.WORKGROUP_SIZE" }
        }
      ]
    }
  ]
}