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 for the reference semantics.

Inputs

Name	Bind key	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, but this kernel ABI requires an explicit tensor; callers representing omission must bind a zero-filled tensor.	required
`sequence_lens`	`sequence_lens`	`int32`	`1`	—	Per-batch sequence lengths of shape `[batch_size]`; assumed `seq_length` for all sequences if absent.	optional
`initial_h`	`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`	`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	Bind key	Logical dtype	Rank	Shape	Description	Presence
`Y`	`y`	`T`	`4`	derived; see description	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; see description	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; see description	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
`layout`	`0`	Tensor layout convention: `0` expects `[seq_length, batch_size, ...]` for `X`/`Y` and `[num_directions, batch_size, hidden_size]` for states; `1` swaps `seq_length` and `batch_size`.
`direction`	`"forward"`	Computation direction: `"forward"`, `"reverse"`, or `"bidirectional"`.
`input_forget`	`0`	When set to `1`, couples the input gate and forget gate (i.e. `ft = 1 - it`).
`hidden_size`	—	Optional number of neurons in the hidden layer; when omitted, it is inferred from the W and R tensor shapes.
`clip`	—	Optional non-negative threshold applied to activation inputs as `[-clip, +clip]`; omission disables clipping, while an explicit `0` clamps them to zero.
`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.

Type constraints

Variable	Allowed dtypes
`T`	`float32`

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
datamove-transpose-2d-tiled.wgsl.jinja
lstm-general.wgsl.jinja
lstm-parallel-global.wgsl.jinja
lstm-recur-step.wgsl.jinja
lstm-recur-subgroup-batched-step.wgsl.jinja
lstm-recur-subgroup-step.wgsl.jinja
lstm-seed-state.wgsl.jinja
lstm-timestep-body.wgsl.jinja
matmul-notrans-vec4-tiled-reg.wgsl.jinja
rnn-input-proj.wgsl.jinja
rnn-sequence-mask.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/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] },
});

Downloads last month: -

kernel

webgpu

wgsl

apache-2.0

WebGPU

Requires WebGPU support. See the compatibility table.