| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.LinearAttentionGate |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 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. |
| |
| See the [ONNX Runtime `LinearAttentionGate` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.LinearAttentionGate) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `a` | `aT` | `T` | — | — | 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. | required | |
| | `dt_bias` | `dtBiasT` | `TF` | `1` | — | Per-head float32 bias added to `a`, with shape (H). | required | |
| | `decay_scale` | `decayScaleT` | `TF` | `1` | — | Per-head float32 multiplier applied to `softplus(a + dt_bias)`, with shape `(H)`. For gated DeltaNet this is `-exp(A_log)`. | required | |
| | `b` | `bT` | `T` | — | — | Update-rate projection with the same shape as `a` when `beta` is requested. It is accepted but unused when `beta` is omitted. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `decay` | `decayT` | `T` | same as `a` | same as `a` | `decay_scale * softplus(a + dt_bias)`, with the same shape as `a`. | required | |
| | `beta` | `betaT` | `T` | same as `a` | same as `a` | sigmoid(b), with the same shape as `a`. Requires the `b` input. | optional | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16` | |
| | `TF` | `float32` | |
|
|
| ## Files |
|
|
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, provenance) |
| - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth) |
| - [`test.json`](build/webgpu/test.json) — correctness cases |
| - [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases |
| - [`linear-attention-gate.wgsl.jinja`](build/webgpu/linear-attention-gate.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| The loader derives every required output's shape and logical dtype from the manifest contract and this call. |
| It then allocates the result tensors automatically. |
|
|
| The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version. |
|
|
| Replace each `*Data` placeholder with a typed array containing the corresponding input data. |
|
|
| ```js |
| import { getKernel } from "@huggingface/kernels"; |
| |
| const kernel = await getKernel("webgpu-kernels/com.microsoft.LinearAttentionGate", { version: 1 }); |
| const { decayT } = await kernel({ |
| aT: { data: aTData, shape: [5] }, |
| dtBiasT: { data: dtBiasTData, shape: [5] }, |
| decayScaleT: { data: decayScaleTData, shape: [5] }, |
| }); |
| ``` |
|
|
