| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.GatedRMSNorm |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Gated RMS normalization as used by Mamba2 and gated DeltaNet attention outputs: `Y = X * rsqrt(mean(X^2) + epsilon) * scale * SiLU(gate)`. The mean of squares is taken over each contiguous group of `C` elements, where `C` is the element count of `scale`, so a per-head norm runs on a packed `(..., H * C)` tensor with no surrounding reshape. Rows are counted flat as the element count of `X` divided by `C` and any rank of at least 1 is accepted. Every intermediate including SiLU is float32 whatever the tensor type; only the store narrows. Bfloat16 is not implemented. |
| |
| See the [ONNX Runtime `GatedRMSNorm` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.GatedRMSNorm) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `xT` | `X` | `T` | — | — | Input activations with shape `(..., H * C)`. Any rank of at least 1 is accepted and the leading axes are folded into the row count. | required | |
| | `scaleT` | `scale` | `T` | — | — | Normalization weight. Its element count is the normalization span `C`; the standard spelling is `(C)`, and the reshaped `(1, C)` and `(C, 1)` spellings of the same vector are accepted because upstream measures the span as an element count. | required | |
| | `gateT` | `gate` | `T` | — | — | Gate activations with exactly the same shape as `X`; SiLU is applied to them internally. | required | |
| |
| ## Outputs |
| |
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `yT` | `Y` | `T` | same as `xT` | same as `xT` | Normalized, scaled and gated activations with the same shape and dtype as `X`. | required | |
| |
| ## Attributes |
| |
| Default values (overridable per request): |
| |
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `epsilon` | `0.00001` | Constant added to the mean of squares before the reciprocal square root. The standard default is `1e-5`; note that the Hugging Face gated-RMS archetype defaults to `1e-6` instead, so a value must be carried explicitly when porting between the two. | |
| |
| ## Type constraints |
| |
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `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. |
| |
| - `subgroup_rows_vec4` — One aligned lane group per normalization row inside a subgroup, with the row held in registers. The fold is a segmented butterfly over the low lane bits, so several rows share a subgroup with no workgroup memory, no barrier and no idle lanes beyond the row's own power-of-two padding. |
| - `packed_rows_vec4` — Portable vec4 route: the workgroup is split into a power-of-two lane group per row and each group folds its own contiguous slice of the shared array, so a short row does not idle the workgroup and no subgroup support is required. It also carries rows too wide to stage in registers, which it walks twice. |
| - `subgroup_rows_scalar` — The barrier-free lane-group schedule for a normalization span that is not a multiple of four: the same segmented butterfly over scalar loads. |
| - `packed_rows_scalar` — Scalar fallback for a span that is not a multiple of four, and the route every device can select. It keeps the lane-group-per-row schedule so a narrow span still fills the workgroup. |
| - `subgroup_wide_rows_vec4` — Subgroup partials for a row spanning several whole subgroups; every lane reads the short row sum after one barrier. |
| - `subgroup_wide_rows_scalar` — Subgroup partials for a row spanning several whole subgroups; every lane reads the short row sum after one barrier. |
|
|
| ## Device requirements |
|
|
| Some implementation variants require `subgroups`. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype. |
|
|
| ## 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 cases |
| - [`gated-rms-norm-rows.wgsl.jinja`](build/webgpu/gated-rms-norm-rows.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.3 |
| ``` |
|
|
| 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.GatedRMSNorm", { version: 1 }); |
| const { yT } = await kernel({ |
| xT: { data: xTData, shape: [3, 4] }, |
| scaleT: { data: scaleTData, shape: [1] }, |
| gateT: { data: gateTData, shape: [3, 4] }, |
| }); |
| ``` |
|
|
