| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.BiasSoftmax |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Computes `softmax(data + bias)` over the flattened suffix beginning at `axis`. The required `is_inner_broadcast` attribute selects how bias rows are reused: consecutive groups for inner broadcast or cyclic groups for outer broadcast. This specializes the `softmax(scores + additive_mask)` pattern used by transformer attention. Float16 and float32 are supported; the schema's double type is not. |
|
|
| See the [ONNX Runtime `BiasSoftmax` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.BiasSoftmax) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `data` | `T` | — | — | The input data tensor. | required | |
| | `bias` | `T` | — | — | The bias (or additive mask) tensor. Its element count must be an integral number of flattened softmax rows and that row count must divide the data row count. | required | |
|
|
| ## Outputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `output` | `T` | same as `data` | same as `data` | The output tensor; same shape as data. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | `1` | The axis from which softmax is applied; dimensions from `axis` onward are included in the softmax reduction. | |
| | `is_inner_broadcast` | — | Let `data_row` and `bias_row` index flattened rows of length `product(data.shape[axis:])`. When 1, consecutive data-row groups reuse each bias row: `bias_row = floor(data_row / (data_row_count / bias_row_count))`. When 0, bias rows repeat cyclically: `bias_row = data_row % bias_row_count`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16` | |
|
|
| ## Implementation variants |
|
|
| One implementation is selected per call from the device capabilities, the request shapes and the dtypes; these notes say what each one covers. |
|
|
| - `online_subgroup_rows_vec4` — Subgroup-per-row fused bias and softmax with register-resident vec4 values, f32 addition and reduction, and the existing zero-output convention for fully masked rows. |
| - `online_workgroup_vec4` — Vectorized workgroup online bias-softmax with f32 addition and reduction; portable fallback below the register-resident subgroup-row path. |
| - `online_workgroup_scalar_vec4` — Logical vec4 online reduction with scalar storage for unaligned row strides. Invalid tail components contribute zero denominator and are never stored. |
|
|
| ## 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, 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 |
| - [`bias-softmax-longrow-normalize.wgsl.jinja`](build/webgpu/bias-softmax-longrow-normalize.wgsl.jinja) |
| - [`bias-softmax.wgsl.jinja`](build/webgpu/bias-softmax.wgsl.jinja) |
| - [`softmax-longrow-stats.wgsl.jinja`](build/webgpu/softmax-longrow-stats.wgsl.jinja) |
| - [`softmax-online.wgsl.jinja`](build/webgpu/softmax-online.wgsl.jinja) |
| - [`softmax-subgroup-rows.wgsl.jinja`](build/webgpu/softmax-subgroup-rows.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/com.microsoft.BiasSoftmax", { version: 1 }); |
| const { output } = await kernel({ |
| data: { data: dataData, shape: [1, 2, 2] }, |
| bias: { data: biasData, shape: [1, 2, 2] }, |
| }, { |
| attrs: { is_inner_broadcast: 1 }, |
| }); |
| ``` |
|
|
