| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.RMSNormalization |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 23 |
|
|
| ## Description |
|
|
| Computes RMS normalization over the suffix dimensions of `X` starting at `axis`: `Y = X / sqrt(mean(X^2) + epsilon) * scale`. The normalization stage supports TensorProto `stash_type` values `1` (float32) and `10` (float16), and is cast back to the dtype of `X` before `scale` is applied. The input type `T` and scale/output type `V` may independently be float16 or float32; ONNX's bfloat16 and double cases are not yet implemented. |
|
|
| See the [ONNX `RMSNormalization` spec](https://onnx.ai/onnx/operators/onnx__RMSNormalization.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `X` | `x` | `T` | — | — | Input tensor to be normalized; the RMS is taken over the last dimensions starting at `axis`. | required | |
| | `scale` | `scale` | `V` | — | — | Scale tensor, unidirectionally broadcastable to `X`; its dtype `V` may differ from the input dtype `T`. | required | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `Y` | `y` | `V` | same as `X` | same as `X` | Normalized and scaled output tensor; same shape as `X` and same dtype `V` as `scale`. | required | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | `-1` | The first dimension of the normalization suffix; negative values count from the end, so the default `-1` normalizes over only the last dimension. | |
| | `epsilon` | `0.00001` | Small constant added to the mean square before taking the square root to avoid division by zero. | |
| | `stash_type` | `1` | TensorProto element type used for normalization: `1` computes in float32, while `10` computes in float16. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16` | |
| | `V` | `float32`, `float16` | |
|
|
| ## 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 |
| - [`norm-row-stats.wgsl.jinja`](build/webgpu/norm-row-stats.wgsl.jinja) |
| - [`rms-normalization-splitk-normalize.wgsl.jinja`](build/webgpu/rms-normalization-splitk-normalize.wgsl.jinja) |
| - [`rms-normalization-splitk-partials.wgsl.jinja`](build/webgpu/rms-normalization-splitk-partials.wgsl.jinja) |
| - [`rms-normalization-stash-f16-serial.wgsl.jinja`](build/webgpu/rms-normalization-stash-f16-serial.wgsl.jinja) |
| - [`rms-normalization.wgsl.jinja`](build/webgpu/rms-normalization.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/ai.onnx.RMSNormalization", { version: 1 }); |
| const { y } = await kernel({ |
| x: { data: xData, shape: [1, 2, 3] }, |
| scale: { data: scaleData, shape: [3] }, |
| }); |
| ``` |
|
|
