| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.MatMulNBits |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Matrix multiplication with `B` block-quantized along K and dequantized as `(code - zero_point) * scale`. Each power-of-two `block_size` group has a scale and optional zero point; optional bias is added afterward. Two-, four-, and eight-bit codes are packed low-first, and `A` may have rank 2 or 3. This package supports standard unpacked zero points with the same dtype as `A`. Deprecated `g_idx`, prepacked weights, and bfloat16 tensors are not implemented. |
|
|
| See the [ONNX Runtime `MatMulNBits` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.MatMulNBits) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `aT` | `A` | `T1` | — | — | Float input matrix, not quantized. Rank 2 has shape `(M, K)` and rank 3 has shape `(batch, sequence, K)`; only the last axis is the reduction axis and the leading axes fold into the row count, so the ordinary activation needs no surrounding Reshape. | required | |
| | `bT` | `B` | `uint8` | `3` | — | Bit-packed uint8 weight matrix of shape `(N, k_blocks, blob_size)`, where `k_blocks = ceil(K / block_size)` and `blob_size = block_size * bits / 8`. Codes are packed low-first along K. Bound in the packed storage layout: four blob bytes per u32 word, so the kernels stream the blob's own bytes rather than one widened word per byte. | required | |
| | `scalesT` | `scales` | `T1` | `2` | — | Per-block dequantization scale factors of shape `(N, k_blocks)`, with the same dtype as `A`. | required | |
| | `zeroPointsT` | `zero_points` | `T3` | `2` | — | Standard unpacked per-block zero points with shape `(N, k_blocks)` and the same dtype as `A`. Omission uses `2^(bits - 1)`. | optional | |
| | `biasT` | `bias` | `T1` | `1` | — | Optional bias vector of shape `[N]` added to the output. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `yT` | `Y` | `T1` | same as `aT` | derived | Result of A multiplied by the dequantized weight matrix, with optional bias, same dtype and rank as A: the leading axes of A with a trailing N. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `K` | — | Input feature dimension of the weight matrix. | |
| | `N` | — | Output feature dimension of the weight matrix. | |
| | `accuracy_level` | `0` | Minimum internal accuracy level: 0 (unset), 1 (float32), 2 (float16), 3 (bfloat16), or 4 (int8). | |
| | `bits` | `4` | Bit width used to quantize B; this package supports 2, 4, and 8. | |
| | `block_size` | — | Power-of-two quantization block size along K; it must be at least 16. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T1` | `float32`, `float16` | |
| | `T3` | `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. |
|
|
| - `prefill_tiled_reg_vec4_splitk_default_zero` — Cuts the K reduction of the four-wide register-blocked prefill tile into power-of-two slices across dispatch.z, retaining at least 512 K values per slice, then sums the f32 partials and adds bias in a second pass. |
| - `prefill_tiled_reg_vec4_default_zero` — The register-blocked prefill tile with four-wide activation loads and a 128-row tile above 256 rows: the activation slice of every K tile is staged with one vector load per lane instead of four scalar loads, and the taller tile halves the dequantization work per multiply-add. |
| - `prefill_tiled_reg_vec4_splitk_zero_bias` — Cuts the K reduction of the four-wide register-blocked prefill tile into power-of-two slices across dispatch.z, retaining at least 512 K values per slice, then sums the f32 partials and adds bias in a second pass. |
| - `prefill_tiled_reg_vec4_zero_bias` — The register-blocked prefill tile with four-wide activation loads and a 128-row tile above 256 rows: the activation slice of every K tile is staged with one vector load per lane instead of four scalar loads, and the taller tile halves the dequantization work per multiply-add. |
| - `prefill_tiled_reg_vec4_splitk_zero_only` — Cuts the K reduction of the four-wide register-blocked prefill tile into power-of-two slices across dispatch.z, retaining at least 512 K values per slice, then sums the f32 partials and adds bias in a second pass. |
| - `prefill_tiled_reg_vec4_zero_only` — The register-blocked prefill tile with four-wide activation loads and a 128-row tile above 256 rows: the activation slice of every K tile is staged with one vector load per lane instead of four scalar loads, and the taller tile halves the dequantization work per multiply-add. |
| - `prefill_tiled_reg_vec4_splitk_bias_only` — Cuts the K reduction of the four-wide register-blocked prefill tile into power-of-two slices across dispatch.z, retaining at least 512 K values per slice, then sums the f32 partials and adds bias in a second pass. |
| - `prefill_tiled_reg_vec4_bias_only` — The register-blocked prefill tile with four-wide activation loads and a 128-row tile above 256 rows: the activation slice of every K tile is staged with one vector load per lane instead of four scalar loads, and the taller tile halves the dequantization work per multiply-add. |
|
|
| ## Device requirements |
|
|
| Some implementation variants require `subgroup-matrix` and `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 + tuning cases |
| - [`matmul-nbits-dp4a-quantize.wgsl.jinja`](build/webgpu/matmul-nbits-dp4a-quantize.wgsl.jinja) |
| - [`matmul-nbits-gemv-q4.wgsl.jinja`](build/webgpu/matmul-nbits-gemv-q4.wgsl.jinja) |
| - [`matmul-nbits-q4-dp4a-prefill.wgsl.jinja`](build/webgpu/matmul-nbits-q4-dp4a-prefill.wgsl.jinja) |
| - [`matmul-nbits-q4-prefill-tile4x4.wgsl.jinja`](build/webgpu/matmul-nbits-q4-prefill-tile4x4.wgsl.jinja) |
| - [`matmul-nbits-q4-prefill-tiled-reg.wgsl.jinja`](build/webgpu/matmul-nbits-q4-prefill-tiled-reg.wgsl.jinja) |
| - [`matmul-nbits-q4-prefill-tiled.wgsl.jinja`](build/webgpu/matmul-nbits-q4-prefill-tiled.wgsl.jinja) |
| - [`matmul-nbits-q4-sgmat.wgsl.jinja`](build/webgpu/matmul-nbits-q4-sgmat.wgsl.jinja) |
| - [`matmul-nbits.wgsl.jinja`](build/webgpu/matmul-nbits.wgsl.jinja) |
| - [`reduce-axis0-splitk-combine.wgsl.jinja`](build/webgpu/reduce-axis0-splitk-combine.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.2 |
| ``` |
|
|
| 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.MatMulNBits", { version: 1 }); |
| const { yT } = await kernel({ |
| aT: { data: aTData, shape: [2, 17] }, |
| bT: { data: bTData, shape: [2, 2, 8] }, |
| scalesT: { data: scalesTData, shape: [2, 2] }, |
| }, { |
| attrs: { K: 17, N: 2, block_size: 16 }, |
| }); |
| ``` |
|
|
