Xenova HF Staff

sync 2e7068faf55e

a44211c verified 5 days ago

2.61 kB

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

com.microsoft.BiasAdd

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

Description

Adds a 1-D bias (broadcast over the channel dimension) to input X, then adds the residual tensor skip elementwise. All three tensors share the same channel count C; X and skip have shape (N, S, C).

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

Inputs

Name	Bind key	Logical dtype	Rank	Shape	Description	Presence
`X`	`X`	`T`	`3`	—	Input tensor of shape `(N, S, C)`: batch size `N`, spatial size `S`, and `C` channels.	required
`bias`	`bias`	`T`	`1`	—	1-D bias vector of length C, broadcast-added along the channel dimension.	required
`skip`	`skip`	`T`	`3`	—	Residual tensor with the same `(N, S, C)` shape as `X`, added after the bias.	required

Outputs

Name	Bind key	Logical dtype	Rank	Shape	Description	Presence
`Y`	`Y`	`T`	`3`	same as `X`	Output tensor of shape `(N, S, C)`: the elementwise sum `X + bias + skip`.	required

Type constraints

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

Files

metadata.json — kernel metadata (id, digests, provenance)
manifest.json — the op contract (source of truth)
test.json — correctness cases
bench.json — benchmark + tuning cases
bias-add.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.

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

const kernel = await getKernel("webgpu-kernels/com.microsoft.BiasAdd", { version: 1 });
const { Y } = await kernel({
  X: { data: XData, shape: [1, 2, 4] },
  bias: { data: biasData, shape: [4] },
  skip: { data: skipData, shape: [1, 2, 4] },
});