Xenova HF Staff

sync 2e7068faf55e

53a7bdc verified 5 days ago

4.38 kB

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

com.microsoft.CausalConvWithState

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

Description

Legacy Microsoft contrib form of stateful 1-D causal depthwise convolution. Each channel uses its own (channels, 1, kernel) weight over current and past positions, with optional activation and past_state/present_state tensors for incremental decoding. The contrib-only state_window attribute may retain several rollback states. This inference implementation preserves the existing contrib ABI with ndim = 1, float16 or float32 tensors, and float32 accumulation; spatial ranks 2 and 3 and bfloat16 are not implemented.

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

Inputs

Name	Bind key	Logical dtype	Rank	Shape	Description	Presence
`input`	`inputT`	`T`	`3`	—	Channels-first input tensor with shape `(batch_size, channels, sequence_length)` for the supported 1-D mode.	required
`weight`	`weightT`	`T`	`3`	—	Depthwise convolution kernel with shape `(channels, 1, kernel_size)` for the supported 1-D mode.	required
`bias`	`biasT`	`T`	`1`	—	Optional per-channel bias with shape `(channels,)`.	optional
`past_state`	`pastStateT`	`T`	derived	—	Carry state from the previous step; shape `(batch_size, channels, k_1 - 1)`, or `(W, batch_size, channels, k_1 - 1)` when `state_window = W > 0`, in which case only slot `W - 1` is read. If absent, the left-side padding is zero.	optional

Outputs

Name	Bind key	Logical dtype	Rank	Shape	Description	Presence
`output`	`outputT`	`T`	`3`	same as `input`	Convolution output with the same shape as `input`.	required
`present_state`	`presentStateT`	`T`	derived	derived; see description	Updated carry state; shape `(batch_size, channels, k_1 - 1)`, or `(W, batch_size, channels, k_1 - 1)` when `state_window = W > 0`. Slot `W - 1` holds the last `k - 1` values along the causal axis; slot `j` holds the same for the prefix ending at position `seq_len - W + j`.	required

Attributes

Default values (overridable per request):

Attribute	Default	Description
`activation`	`"none"`	Activation applied after convolution and bias. Defaults to `none`; `swish` is an alias of SiLU.
`ndim`	`1`	Number of spatial dimensions. This implementation supports the contrib 1D mode (`ndim = 1`).
`state_window`	`0`	Contrib extension selecting the number of rollback state slots to retain, in the range 0 through 8. Defaults to 0.

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
causal-conv-with-state-tiled.wgsl.jinja
causal-conv-with-state-vec4.wgsl.jinja
causal-conv-with-state.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.CausalConvWithState", { version: 1 });
const { outputT, presentStateT } = await kernel({
  inputT: { data: inputTData, shape: [1, 1, 5] },
  weightT: { data: weightTData, shape: [1, 1, 4] },
});