| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.LinearAttention |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Recurrent linear attention for packed `[B, T, H*D]` decode and prefill. It supports all four update rules, standard and inverse GQA, shared-key heads, and rollback states through `state_window`. Activations and state may independently use float16 or float32; bfloat16 is not implemented. `past_state` is optional for every update rule and defaults to zeros. |
| |
| See the [ONNX Runtime `LinearAttention` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.LinearAttention) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `queryT` | `query` | `T` | `3` | — | Query vectors with 3D packed shape `(B, T, H_q * d_k)`; heads are packed into the last dimension. | required | |
| | `keyT` | `key` | `T` | `3` | — | Key vectors with 3D packed shape `(B, T, H_k * d_k)`, where positive `H_k` divides `H_kv`; `H_k < H_kv` shares each key head across multiple KV-state heads. Keys should be L2-normalized for `delta`/`gated_delta` modes. | required | |
| | `valueT` | `value` | `T` | `3` | — | Value vectors with 3D packed shape `(B, T, H_kv * d_v)`. | required | |
| | `pastStateT` | `past_state` | `S` | derived | derived | Recurrent state from the previous step with shape `(B, H_kv, d_k, d_v)`, or `(W, B, H_kv, d_k, d_v)` when `state_window = W > 0`; defaults to zeros if absent. | optional | |
| | `decayT` | `decay` | `T` | `3` | — | Exponential decay gate in log-space with shape `(B, T, H_kv * d_k)` or `(B, T, H_kv)`; required for `gated` and `gated_delta` modes. | optional | |
| | `betaT` | `beta` | `T` | `3` | — | Update rate (sigmoid output) with shape `(B, T, H_kv)` or `(B, T, 1)`; required for `delta` and `gated_delta` modes. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `outputT` | `output` | `T` | `3` | derived | Attention output with 3D packed shape `(B, T, max(H_q, H_kv) * d_v)`. | required | |
| | `presentStateT` | `present_state` | `S` | derived | derived | Updated recurrent state with shape `(B, H_kv, d_k, d_v)`, or `(W, B, H_kv, d_k, d_v)` when `state_window = W > 0`. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `chunk_size` | `64` | Accepted for schema compatibility; does not affect the result. | |
| | `kv_num_heads` | — | Number of key/value heads. | |
| | `q_num_heads` | — | Number of query heads. | |
| | `scale` | `0` | Scale applied to query-key products. Zero selects `1 / sqrt(d_k)`. | |
| | `state_window` | `0` | Number of recent recurrent states retained in `present_state`, in the supported range 0 to 8; zero returns only the current state. | |
| | `update_rule` | `"gated_delta"` | Recurrent update rule: `linear`, `gated`, `delta`, or `gated_delta`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16` | |
| | `S` | `float32`, `float16` | |
|
|
| ## Implementation variants |
|
|
| One implementation is selected per call from the device capabilities, the request shapes and the dtypes; these notes say what each one covers. |
|
|
| - `linear_zero_serial_small_dk` — Small key dimensions keep each state column private across the sequence. Linear updates specialize shape-known indexing and loop bounds, while gated-delta updates retain their uniform-driven recurrence. |
| - `linear_state_serial_small_dk` — Small key dimensions keep each state column private across the sequence. Linear updates specialize shape-known indexing and loop bounds, while gated-delta updates retain their uniform-driven recurrence. |
| - `gated_delta_zero_serial_small_dk` — Small key dimensions keep each state column private across the sequence. Linear updates specialize shape-known indexing and loop bounds, while gated-delta updates retain their uniform-driven recurrence. |
| - `gated_delta_state_serial_small_dk` — Small key dimensions keep each state column private across the sequence. Linear updates specialize shape-known indexing and loop bounds, while gated-delta updates retain their uniform-driven recurrence. |
|
|
| ## Files |
|
|
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, 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 |
| - [`chunk-out.wgsl.jinja`](build/webgpu/chunk-out.wgsl.jinja) |
| - [`chunk-prep.wgsl.jinja`](build/webgpu/chunk-prep.wgsl.jinja) |
| - [`chunk-scan.wgsl.jinja`](build/webgpu/chunk-scan.wgsl.jinja) |
| - [`chunk-ut.wgsl.jinja`](build/webgpu/chunk-ut.wgsl.jinja) |
| - [`linear-attention.scalar.wgsl.jinja`](build/webgpu/linear-attention.scalar.wgsl.jinja) |
| - [`linear-attention.serial.wgsl.jinja`](build/webgpu/linear-attention.serial.wgsl.jinja) |
| - [`linear-attention.vec4.wgsl.jinja`](build/webgpu/linear-attention.vec4.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.2 |
| ``` |
|
|
| Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically. |
|
|
| The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version. |
| It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `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.LinearAttention", { version: 1 }); |
| const { outputT, presentStateT } = await kernel({ |
| queryT: { data: queryTData, shape: [1, 3, 8] }, |
| keyT: { data: keyTData, shape: [1, 3, 4] }, |
| valueT: { data: valueTData, shape: [1, 3, 4] }, |
| pastStateT: { data: pastStateTData, shape: [1, 1, 4, 4] }, |
| decayT: { data: decayTData, shape: [1, 3, 1] }, |
| betaT: { data: betaTData, shape: [1, 3, 1] }, |
| }, { |
| attrs: { q_num_heads: 2, kv_num_heads: 1 }, |
| }); |
| ``` |
|
|
