| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.Concat |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 13 |
|
|
| ## Description |
|
|
| Concatenates a list of tensors along a single axis into one output tensor. All inputs must have the same shape except along the concatenation `axis`, which may differ between inputs. Negative `axis` values count from the end of the rank. |
|
|
| See the [ONNX `Concat` spec](https://onnx.ai/onnx/operators/onnx__Concat.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `a` | `A` | `T` | — | — | First input tensor to concatenate. | required | |
| | `b` | `B` | `T` | — | — | Second input tensor to concatenate (optional). | optional | |
| | `d` | `D` | `T` | — | — | Third input tensor to concatenate (optional). | optional | |
| | `e` | `E` | `T` | — | — | Fourth input tensor to concatenate (optional). | optional | |
| | `f` | `F` | `T` | — | — | Fifth input tensor to concatenate (optional). | optional | |
| | `g` | `G` | `T` | — | — | Sixth input tensor to concatenate (optional). | optional | |
| | `h` | `H` | `T` | — | — | Seventh input tensor to concatenate (optional). | optional | |
| | `i` | `I` | `T` | — | — | Eighth input tensor to concatenate (optional). | optional | |
| | `j` | `J` | `T` | — | — | Ninth input tensor to concatenate (optional). | optional | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `c` | `concat_result` | `T` | same as `a` | derived | Concatenated output tensor. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | — | Required axis along which to concatenate; negative values count from the back, and the accepted range is `[-r, r-1]` where `r` is the rank of the inputs. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16`, `uint32`, `int32`, `int16`, `uint8`, `int8`, `bool`, `int64` | |
|
|
| ## 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 |
| - [`concat-last-axis-vec4-tail.wgsl.jinja`](build/webgpu/concat-last-axis-vec4-tail.wgsl.jinja) |
| - [`concat-n.wgsl.jinja`](build/webgpu/concat-n.wgsl.jinja) |
| - [`datamove-concat-block.wgsl.jinja`](build/webgpu/datamove-concat-block.wgsl.jinja) |
| - [`datamove-concat-last-axis-cooperative.wgsl.jinja`](build/webgpu/datamove-concat-last-axis-cooperative.wgsl.jinja) |
| - [`datamove-flat-copy.wgsl.jinja`](build/webgpu/datamove-flat-copy.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.Concat", { version: 1 }); |
| const { c } = await kernel({ a: { data: aData, shape: [1] } }, { |
| attrs: { axis: 0 }, |
| }); |
| ``` |
|
|
