| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.LSTM |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 22 |
|
|
| ## Description |
|
|
| Computes a single-layer LSTM over an input sequence using input, output, forget, and cell gates at each time step. Supports `forward`, `reverse`, and `bidirectional` directions, optional peephole connections, and optional per-sequence length masking. |
|
|
| See the [ONNX `LSTM` spec](https://onnx.ai/onnx/operators/onnx__LSTM.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `x` | `X` | `T` | `3` | — | Input sequences with shape `[seq_length, batch_size, input_size]` when `layout=0`, or `[batch_size, seq_length, input_size]` when `layout=1`. | required | |
| | `w` | `W` | `T` | `3` | — | Gate weight tensor of shape `[num_directions, 4*hidden_size, input_size]`, concatenating weights for input, output, forget, and cell gates. | required | |
| | `r` | `R` | `T` | `3` | — | Recurrence weight tensor of shape `[num_directions, 4*hidden_size, hidden_size]`, concatenating recurrence weights for all gates. | required | |
| | `b` | `B` | `T` | `2` | — | Bias tensor of shape `[num_directions, 8*hidden_size]`, concatenating input and recurrence biases for all gates. ONNX defines an omitted bias as zero; this package requires an explicit tensor, which may be zero-filled to represent omission. | required | |
| | `sequence_lens` | — | `int32` | `1` | — | Per-batch sequence lengths of shape `[batch_size]`; assumed `seq_length` for all sequences if absent. | optional | |
| | `initial_h` | — | `T` | `3` | — | Initial hidden state with shape `[num_directions, batch_size, hidden_size]` when `layout=0`, or `[batch_size, num_directions, hidden_size]` when `layout=1`; assumed zero if absent. | optional | |
| | `initial_c` | — | `T` | `3` | — | Initial cell state with shape `[num_directions, batch_size, hidden_size]` when `layout=0`, or `[batch_size, num_directions, hidden_size]` when `layout=1`; assumed zero if absent. | optional | |
| | `p` | `P` | `T` | `2` | — | Peephole weight tensor of shape `[num_directions, 3*hidden_size]` for input, output, and forget gates; assumed zero if absent. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `y` | `Y` | `T` | `4` | derived | Intermediate hidden states for all time steps, with shape `[seq_length, num_directions, batch_size, hidden_size]` when `layout=0`, or `[batch_size, seq_length, num_directions, hidden_size]` when `layout=1`. | required | |
| | `y_h` | `Y_h` | `T` | `3` | derived | Final hidden state with shape `[num_directions, batch_size, hidden_size]` when `layout=0`, or `[batch_size, num_directions, hidden_size]` when `layout=1`. | required | |
| | `y_c` | `Y_c` | `T` | `3` | derived | Final cell state with shape `[num_directions, batch_size, hidden_size]` when `layout=0`, or `[batch_size, num_directions, hidden_size]` when `layout=1`. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `activation_alpha` | — | Optional alpha parameters for activation functions that use alpha, consumed in activation-list order; omitted entries use the ONNX defaults for their activation. | |
| | `activation_beta` | — | Optional beta parameters for activation functions that use beta, consumed in activation-list order; omitted entries use the ONNX defaults for their activation. | |
| | `activations` | — | Activation functions for the input/output/forget gates, cell candidate, and cell output. Defaults to `["Sigmoid", "Tanh", "Tanh"]` per direction. | |
| | `clip` | — | Optional non-negative threshold applied to activation inputs as `[-clip, +clip]`; omission disables clipping, while an explicit `0` clamps them to zero. | |
| | `direction` | `"forward"` | Computation direction: `"forward"`, `"reverse"`, or `"bidirectional"`. | |
| | `hidden_size` | — | Optional number of neurons in the hidden layer; when omitted, it is inferred from the W and R tensor shapes. | |
| | `input_forget` | `0` | When set to `1`, couples the input gate and forget gate (i.e. `ft = 1 - it`). | |
| | `layout` | `0` | Tensor layout convention: `0` uses `X=[seq_length, batch_size, input_size]`, `Y=[seq_length, num_directions, batch_size, hidden_size]`, and states `[num_directions, batch_size, hidden_size]`; `1` uses `X=[batch_size, seq_length, input_size]`, `Y=[batch_size, seq_length, num_directions, hidden_size]`, and states `[batch_size, num_directions, hidden_size]`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32` | |
|
|
| ## Files |
|
|
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, 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 cases |
| - [`datamove-transpose-2d-tiled.wgsl.jinja`](build/webgpu/datamove-transpose-2d-tiled.wgsl.jinja) |
| - [`lstm-general.wgsl.jinja`](build/webgpu/lstm-general.wgsl.jinja) |
| - [`lstm-parallel-global.wgsl.jinja`](build/webgpu/lstm-parallel-global.wgsl.jinja) |
| - [`lstm-recur-step.wgsl.jinja`](build/webgpu/lstm-recur-step.wgsl.jinja) |
| - [`lstm-recur-subgroup-batched-step.wgsl.jinja`](build/webgpu/lstm-recur-subgroup-batched-step.wgsl.jinja) |
| - [`lstm-recur-subgroup-step.wgsl.jinja`](build/webgpu/lstm-recur-subgroup-step.wgsl.jinja) |
| - [`lstm-seed-state.wgsl.jinja`](build/webgpu/lstm-seed-state.wgsl.jinja) |
| - [`lstm-timestep-body.wgsl.jinja`](build/webgpu/lstm-timestep-body.wgsl.jinja) |
| - [`matmul-notrans-vec4-tiled-reg.wgsl.jinja`](build/webgpu/matmul-notrans-vec4-tiled-reg.wgsl.jinja) |
| - [`rnn-input-proj.wgsl.jinja`](build/webgpu/rnn-input-proj.wgsl.jinja) |
| - [`rnn-sequence-mask.wgsl.jinja`](build/webgpu/rnn-sequence-mask.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| 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. |
|
|
| ```js |
| import { getKernel } from "@huggingface/kernels"; |
| |
| const kernel = await getKernel("webgpu-kernels/ai.onnx.LSTM", { version: 1 }); |
| const { y, y_h, y_c } = await kernel({ |
| x: { data: xData, shape: [1, 1, 1] }, |
| w: { data: wData, shape: [1, 4, 1] }, |
| r: { data: rData, shape: [1, 4, 1] }, |
| b: { data: bData, shape: [1, 8] }, |
| }); |
| ``` |
|
|
