---
library_name: kernels
license: apache-2.0
tags:
- kernel
- webgpu
- wgsl
---
# ai.onnx.Softmax

`ai.onnx`  ·  standard ONNX operator  ·  ONNX opset ≥ 13

## Description

Computes the normalized exponential (softmax) of the input along a single axis: `exp(x) / sum(exp(x))` reduced over that axis. For finite rows, the result has the same shape as the input, with values in `[0, 1]` that sum to 1 along the softmax axis. The implementation uses max subtraction for numerical stability while preserving the literal ONNX result: a row containing only negative infinity produces NaN values.

See the [ONNX `Softmax` spec](https://onnx.ai/onnx/operators/onnx__Softmax.html) for the reference semantics.

## Inputs

| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- |
| `input` | `x` | `T` | — | — | The input tensor of any shape. | required |

## Outputs

| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- |
| `output` | `y` | `T` | same as `input` | same as `input` | The softmax values; same shape as the input. | required |

## Attributes

Default values (overridable per request):

| Attribute | Default | Description |
| --- | --- | --- |
| `axis` | `-1` | The axis along which softmax is computed. Negative values count from the end; the default `-1` softmaxes over the last dimension. |

## Type constraints

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

## Device requirements

Some implementation variants require `subgroups`. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype.

## 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
- [`softmax-longrow-normalize.wgsl.jinja`](build/webgpu/softmax-longrow-normalize.wgsl.jinja)
- [`softmax-longrow-stats.wgsl.jinja`](build/webgpu/softmax-longrow-stats.wgsl.jinja)
- [`softmax-normalize.wgsl.jinja`](build/webgpu/softmax-normalize.wgsl.jinja)
- [`softmax-online-local.wgsl.jinja`](build/webgpu/softmax-online-local.wgsl.jinja)
- [`softmax-online-packed-rows.wgsl.jinja`](build/webgpu/softmax-online-packed-rows.wgsl.jinja)
- [`softmax-online.wgsl.jinja`](build/webgpu/softmax-online.wgsl.jinja)
- [`softmax-row-stage.wgsl.jinja`](build/webgpu/softmax-row-stage.wgsl.jinja)
- [`softmax-strided-online-coop.wgsl.jinja`](build/webgpu/softmax-strided-online-coop.wgsl.jinja)
- [`softmax-strided-online-lane.wgsl.jinja`](build/webgpu/softmax-strided-online-lane.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/ai.onnx.Softmax", { version: 1 });
const { y } = await kernel({ x: { data: xData, shape: [1, 3] } });
```