| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.QuantizeLinear |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 25 |
|
|
| ## Description |
|
|
| Linearly quantizes a high-precision tensor to a lower-precision integer type using the formula `y = saturate((x / y_scale) + y_zero_point)`, with rounding to nearest even. Supports per-tensor, per-axis, and blocked quantization granularities determined by the shape of `y_scale`. |
|
|
| See the [ONNX `QuantizeLinear` spec](https://onnx.ai/onnx/operators/onnx__QuantizeLinear.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `x` | `x` | `TX` | — | — | N-D full-precision input tensor to be quantized. | required | |
| | `y_scale` | `y_scale` | `TS` | — | — | Scale factor; scalar for per-tensor, 1-D for per-axis, or same rank as `x` (with one axis blocked) for blocked quantization. | required | |
| | `y_zero_point` | `y_zero_point` | `TQ` | — | — | Zero point for quantization; must have the same shape as `y_scale`. Defaults to zero if omitted. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `y` | `y` | `TQ` | same as `x` | same as `x` | N-D quantized output tensor with the same shape as `x`. | required | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | `1` | Axis of the quantization dimension in `x`, used for per-axis and blocked quantization; negative values count from the end. | |
| | `block_size` | `0` | Number of elements along `axis` that share a single scale value for blocked quantization; 0 means blocked quantization is not used. | |
| | `output_dtype` | `0` | ONNX TensorProto element-type code for `y`; 0 infers the type from `y_zero_point`, or uint8 when the zero point is omitted. | |
| | `precision` | `0` | ONNX TensorProto element-type code used for `x / y_scale`; `0` uses the dtype of `y_scale`, `1` selects FLOAT, and `10` selects FLOAT16. | |
| | `saturate` | `1` | Controls out-of-range conversion for float8 outputs. The implemented int8/uint8 subset accepts the ONNX default `1`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `TX` | `float32`, `float16` | |
| | `TS` | `float32`, `float16` | |
| | `TQ` | `uint8`, `int8` | |
|
|
| ## 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 |
| - [`quant-linear-blocked-axis.wgsl.jinja`](build/webgpu/quant-linear-blocked-axis.wgsl.jinja) |
| - [`quant-linear-scalar.wgsl.jinja`](build/webgpu/quant-linear-scalar.wgsl.jinja) |
| - [`quant-linear-vec4.wgsl.jinja`](build/webgpu/quant-linear-vec4.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| The loader automatically allocates outputs whose metadata it can derive from the manifest contract and this call. |
|
|
| The explicit `outputs` entries provide shape and logical dtype metadata for the results listed below: |
|
|
| - `y` |
|
|
| Each entry either requests an optional result or supplies metadata that cannot be inferred from the inputs. |
|
|
| 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.QuantizeLinear", { version: 1 }); |
| // Explicit destinations request optional results or supply metadata that cannot be inferred. |
| const { y } = await kernel({ x: { data: xData, shape: [] }, y_scale: { data: y_scaleData, shape: [] } }, { |
| outputs: { y: { shape: [], dtype: "uint8" } }, |
| }); |
| ``` |
|
|
