Xenova HF Staff

sync 91d990483a17

62736d3 verified 3 days ago

4.37 kB

library_name: kernels
license: apache-2.0
tags:
  - kernel
  - webgpu
  - wgsl

com.microsoft.SkipLayerNormalization

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

Description

Fuses skip addition with layer normalization for rank-2 or rank-3 input and a non-empty hidden axis. With exact-shape skip, float32 and float16 output-only paths support optional beta; adding bias requires beta. Returning the residual sum requires beta: float32 supports optional bias and arbitrary hidden sizes, while float16 requires bias and a hidden size divisible by four. Broadcast skip is supported for rank-3 float32 input, required beta, no bias or residual output, and a hidden size divisible by four. Bfloat16 and training statistics are not implemented.

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

Inputs

Name	Upstream name	Logical dtype	Rank	Shape	Description	Presence
`inputT`	`input`	`T`	—	—	Primary input normalized over the final hidden-size axis. Rank 3 is the standard shape; rank 2 is a supported extension.	required
`skipT`	`skip`	`T`	—	—	Residual tensor. For rank-3 input it is exact shape, `(1, sequence_length, hidden_size)`, or `(sequence_length, hidden_size)`; rank-2 input requires exact shape.	required
`gammaT`	`gamma`	`T`	`1`	—	Layer-norm scale weights of shape `(hidden_size)`.	required
`betaT`	`beta`	`T`	`1`	—	Layer-norm bias weights of shape `(hidden_size)`.	optional
`biasT`	`bias`	`T`	`1`	—	Optional additive bias of shape `(hidden_size)` added to `input + skip` before normalization.	optional

Outputs

Name	Upstream name	Logical dtype	Rank	Shape	Description	Presence
`outputT`	`output`	`T`	same as `inputT`	same as `inputT`	Normalized output tensor with the same shape as `input`.	required
`residualT`	`input_skip_bias_sum`	`T`	same as `inputT`	same as `inputT`	Sum of `input`, `skip`, and `bias` (when present) before normalization, with the same shape as `input`.	optional

Attributes

Default values (overridable per request):

Attribute	Default	Description
`epsilon`	`9.999999960041972e-13`	Non-negative epsilon added to the variance before taking the square root.

Type constraints

Variable	Allowed dtypes
`T`	`float32`, `float16`

Device requirements

Some implementation variants require shader-f16. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype.

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 + tuning cases
norm-skip-row-vec4.wgsl.jinja
norm-skip-row.wgsl.jinja

Use with `@huggingface/kernels`

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.

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

const kernel = await getKernel("webgpu-kernels/com.microsoft.SkipLayerNormalization", { version: 1 });
const { outputT } = await kernel({
  inputT: { data: inputTData, shape: [2, 4] },
  skipT: { data: skipTData, shape: [2, 4] },
  gammaT: { data: gammaTData, shape: [4] },
});