com.microsoft.GroupNorm

com.microsoft · ONNX Runtime contrib operator · contrib since_version 1

Description

Group normalization over a 4-D image tensor with an optional fused SiLU: Y = gamma * (X - mean) / sqrt(variance + epsilon) + beta, with mean and variance taken per batch item over each group of C / groups channels and every spatial position. Both layouts are implemented: channels_last = 1 reads (N, H, W, C) and channels_last = 0 reads (N, C, H, W). X/Y and gamma/beta carry independent element types, so float16 activations with float32 weights is a first-class case. Statistics and the affine are float32 whatever the tensor type; only the store narrows.

See the ONNX Runtime GroupNorm contrib-operator spec for the reference semantics.

Inputs

Outputs

Attributes

Attributes and default values (overridable per request):

Type constraints

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.

• split — Three dispatches: partial group statistics, combine mean and reciprocal standard deviation, then normalize and apply affine weights. Channels-last statistics use four, two, or one aligned values per lane; group blocks and spatial rows follow device limits.
• fused — One workgroup per batch item and group, walking the group twice inside a single dispatch: once for the statistics and once to write the affine result. It avoids the split route's two extra launches on small tensors and is the route that needs no scratch buffer.

Files

• metadata.json — kernel metadata (id, digests, per-variant templates, provenance)
• manifest.json — the op contract (source of truth)
• test.json — correctness cases
• bench.json — benchmark cases
• group-norm-combine.wgsl.jinja
• group-norm-fused.wgsl.jinja
• group-norm-slab.wgsl.jinja

Use with @huggingface/kernels

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.

import { getKernel } from "@huggingface/kernels";

const kernel = await getKernel("webgpu-kernels/com.microsoft.GroupNorm", { version: 1 });
const { Y } = await kernel({
  X: { data: XData, shape: [1, 2, 2, 16] },
  gamma: { data: gammaData, shape: [16] },
  beta: { data: betaData, shape: [16] },
}, {
  attrs: { groups: 4, activation: 0 },
});