| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.ArgMax |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 13 |
|
|
| ## Description |
|
|
| Returns the index of the maximum value along an axis, choosing the first equal value unless `select_last_index` is enabled. |
|
|
| See the [ONNX `ArgMax` spec](https://onnx.ai/onnx/operators/onnx__ArgMax.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `data` | `x` | `T` | — | — | Values whose maximum index is selected along `axis`. | required | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | --- | |
| | `reduced` | `y` | `I` | `uint32` | derived | derived; see description | Logical int64 indices of the maximum values along the reduced axis; WebGPU stores these bounded indices as uint32. | required | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | `0` | Axis to reduce; negative values count from the back. | |
| | `keepdims` | `1` | Retain the reduced dimension with length one when non-zero. | |
| | `select_last_index` | `0` | Choose the last equal maximum instead of the first when non-zero. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16`, `int32`, `uint32`, `int16`, `int8`, `uint8` | |
| | `I` | `int64` | |
|
|
| ## 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 |
| - [`reduce-arg-axis-split-combine.wgsl.jinja`](build/webgpu/reduce-arg-axis-split-combine.wgsl.jinja) |
| - [`reduce-arg-axis-split-reduce.wgsl.jinja`](build/webgpu/reduce-arg-axis-split-reduce.wgsl.jinja) |
| - [`reduce-arg-axis-split-tiled.wgsl.jinja`](build/webgpu/reduce-arg-axis-split-tiled.wgsl.jinja) |
| - [`reduce-arg-axis-tiled.wgsl.jinja`](build/webgpu/reduce-arg-axis-tiled.wgsl.jinja) |
| - [`reduce-arg-axis.wgsl.jinja`](build/webgpu/reduce-arg-axis.wgsl.jinja) |
| - [`reduce-arg-row-split.wgsl.jinja`](build/webgpu/reduce-arg-row-split.wgsl.jinja) |
| - [`reduce-arg-row-subgroup.wgsl.jinja`](build/webgpu/reduce-arg-row-subgroup.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.ArgMax", { version: 1 }); |
| const { y } = await kernel({ x: { data: xData, shape: [2, 2] } }); |
| ``` |
|
|
