| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.GroupNormalization |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 21 |
|
|
| ## Description |
|
|
| Applies group normalization to the input: `y = scale * (x - mean) / sqrt(variance + epsilon) + bias`, where mean and variance are computed per instance per group of channels. The number of groups `num_groups` must divide the channel count `C` evenly; when `num_groups == C` this is equivalent to InstanceNormalization, and when `num_groups == 1` it is equivalent to LayerNormalization. The normalization stage supports TensorProto `stash_type` values `1` (float32) and `10` (float16). |
|
|
| See the [ONNX `GroupNormalization` spec](https://onnx.ai/onnx/operators/onnx__GroupNormalization.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `X` | `x` | `T` | — | — | Input data tensor of shape `(N x C x D1 x ... x Dn)` where `N` is batch size and `C` is the number of channels. | required | |
| | `scale` | `scale` | `T` | `1` | — | Scale tensor of shape `(C)`, one value per channel. | required | |
| | `bias` | `bias` | `T` | `1` | — | Bias tensor of shape `(C)`, one value per channel. | required | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `Y` | `y` | `T` | same as `X` | same as `X` | Normalized output tensor of the same shape as `X`. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `epsilon` | `0.00001` | Small value added to the variance denominator to avoid division by zero. | |
| | `stash_type` | `1` | TensorProto element type used for the normalization stage: `1` computes in float32, while `10` computes in float16. Normalized values are cast back to the input type before scale and bias are applied. | |
| | `num_groups` | — | Required number of groups to divide the channels into; must be a divisor of `C`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `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 |
| - [`group-normalization-splitk-apply.wgsl.jinja`](build/webgpu/group-normalization-splitk-apply.wgsl.jinja) |
| - [`group-normalization-splitk-partials.wgsl.jinja`](build/webgpu/group-normalization-splitk-partials.wgsl.jinja) |
| - [`group-normalization-stash-f16-serial.wgsl.jinja`](build/webgpu/group-normalization-stash-f16-serial.wgsl.jinja) |
| - [`norm-row-stats.wgsl.jinja`](build/webgpu/norm-row-stats.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.GroupNormalization", { version: 1 }); |
| const { y } = await kernel({ |
| x: { data: xData, shape: [1, 2, 3] }, |
| scale: { data: scaleData, shape: [2] }, |
| bias: { data: biasData, shape: [2] }, |
| }, { |
| attrs: { num_groups: 1 }, |
| }); |
| ``` |
|
|
