| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.GatherBlockQuantized |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Gathers rows from a block-wise quantized weight matrix and dequantizes them. This inference implementation supports the standard `gather_axis = 0`, `quantize_axis = 1` matrix subset with uint8 `data`, 4-bit packed or 8-bit values, rank-1 non-negative in-bounds int64 `indices` projected to uint32 WebGPU storage, and float32 scales/output. Higher-rank gathers, negative indices, int32 indices, int4/uint4 data, 2-bit data, float16/bfloat16 output, and non-default axes are not implemented. |
| |
| See the [ONNX Runtime `GatherBlockQuantized` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.GatherBlockQuantized) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Bind key | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | --- | |
| | `data` | `dataT` | `T1` | runtime-selected; narrow integers and bool use 32-bit slots | `2` | — | Constant uint8 weight matrix. With `bits = 4`, each byte stores two values low-nibble first; with `bits = 8`, each byte stores one value. | required | |
| | `indices` | `indicesT` | `Tind` | `uint32` | `1` | — | Non-negative logical int64 indices selecting rows from axis 0 of `data`. Every index must be less than the row count; values use checked uint32 WebGPU storage. | required | |
| | `scales` | `scalesT` | `T2` | same as logical dtype | `2` | — | Per-block dequantization scale factors of shape `(rows, ceil(output_columns / block_size))`. | required | |
| | `zero_points` | `zeroPointsT` | `T1` | runtime-selected; narrow integers and bool use 32-bit slots | `2` | — | Optional uint8 zero points. At 4 bits two zero points are packed per byte along the quantized axis, low-nibble first; at 8 bits the shape matches `scales`. If absent, uint8 data uses 2^(bits-1). | optional | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `output` | `outputT` | `T2` | `2` | derived; see description | Dequantized floating-point output rows corresponding to the gathered indices. | required | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `bits` | `4` | Bits per quantized value. The schema default is 4; this implementation supports 4 or 8. | |
| | `block_size` | `128` | Number of values sharing a scale. Defaults to 128 and must be a power of two at least 16. | |
| | `gather_axis` | `0` | Axis from which values are gathered. This matrix implementation supports the standard default, axis 0. | |
| | `quantize_axis` | `1` | Axis split into quantization blocks. This matrix implementation supports the standard default, axis 1. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T1` | `uint8` | |
| | `T2` | `float32` | |
| | `Tind` | `int64` | |
|
|
| ## 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 |
| - [`gather-block-quantized-q4-pair.wgsl.jinja`](build/webgpu/gather-block-quantized-q4-pair.wgsl.jinja) |
| - [`gather-block-quantized-q8-vec4.wgsl.jinja`](build/webgpu/gather-block-quantized-q8-vec4.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/com.microsoft.GatherBlockQuantized", { version: 1 }); |
| const { outputT } = await kernel({ |
| dataT: { data: dataTData, shape: [4, 8] }, |
| indicesT: { data: indicesTData, shape: [2] }, |
| scalesT: { data: scalesTData, shape: [4, 1] }, |
| }); |
| ``` |
|
|
