| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.VarlenCausalConvWithState |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Stateful causal depthwise convolution over packed token-major variable-length sequences, without reads across sequence boundaries. `initial_state` carries preceding raw samples and `final_state` is fully written. At positive `state_update_capacity`, `capture_count` selects a clamped prefix of raw input tokens for compact `state_update`; inactive slots are zero. SiLU and Swish are aliases. This implementation supports float16 and float32 with float32 accumulation; bfloat16 is not implemented. |
| |
| See the [ONNX Runtime `VarlenCausalConvWithState` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.VarlenCausalConvWithState) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Bind key | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | --- | |
| | `input` | `inputT` | `T` | same as logical dtype | `2` | — | Token-major packed input with shape `(total_tokens, channels)`. | required | |
| | `weight` | `weightT` | `T` | same as logical dtype | `3` | — | Depthwise kernel with shape `(channels, 1, kernel_size)`. | required | |
| | `cumulative_sequence_length` | `cumulativeSequenceLengthT` | `M` | `int32` | `1` | — | Exclusive prefix sums with shape `(batch_size + 1)`; sequence `i` owns tokens `[cum[i], cum[i + 1])`. | required | |
| | `bias` | `biasT` | `T` | same as logical dtype | `1` | — | Optional per-channel bias with shape `(channels,)`. In an ONNX graph an omitted bias must still occupy input index 3 as an empty name so `initial_state` stays at index 4. | optional | |
| | `initial_state` | `initialStateT` | `T` | same as logical dtype | `3` | — | Required committed carry state with shape `(batch_size, channels, kernel_size - 1)`, holding the raw samples immediately preceding this call. | required | |
| | `capture_count` | `captureCountT` | `M` | `int32` | `1` | — | Optional int32 vector with shape `(batch_size)`. Required exactly when `state_update_capacity` is positive; each value is clamped to `[0, min(state_update_capacity, sequence_length)]`. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `output` | `outputT` | `T` | same as `input` | same as `input` | Convolution output with the same shape as `input`. | required | |
| | `final_state` | `finalStateT` | `T` | `3` | derived; see description | State after each sequence's final token, shape `(batch_size, channels, kernel_size - 1)`. Always fully written. | required | |
| | `state_update` | `stateUpdateT` | `T` | `3` | derived; see description | Optional compact transition values with shape `(batch_size, state_update_capacity, channels)`. Active slots contain the original local input tokens and all other slots are zero. | optional | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `activation` | `"none"` | Fused activation applied after convolution and bias. One of `none`, `silu`, or `swish`; the standard default is `none`. | |
| | `state_update_capacity` | `0` | Static number of compact per-request prefix transition values to expose, in `[0, 8]`. The standard default is 0. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16` | |
| | `M` | `int32` | |
|
|
| ## 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 |
| - [`varlen-causal-conv-stream.wgsl.jinja`](build/webgpu/varlen-causal-conv-stream.wgsl.jinja) |
| - [`varlen-causal-conv.wgsl.jinja`](build/webgpu/varlen-causal-conv.wgsl.jinja) |
| - [`varlen-state-update.wgsl.jinja`](build/webgpu/varlen-state-update.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/com.microsoft.VarlenCausalConvWithState", { version: 1 }); |
| const { outputT, finalStateT } = await kernel({ |
| inputT: { data: inputTData, shape: [5, 6] }, |
| weightT: { data: weightTData, shape: [6, 1, 4] }, |
| cumulativeSequenceLengthT: { data: cumulativeSequenceLengthTData, shape: [3] }, |
| initialStateT: { data: initialStateTData, shape: [2, 6, 3] }, |
| }); |
| ``` |
|
|
