| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.MatMulNBitsQkv |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Fuses RMS normalization with three block-quantized attention projections: `A_norm = RMSNorm(A + skip, norm_scale)` (or without `skip`), followed by Q, K, and V projections. The optional fourth output returns `A + skip`. Only 4-bit weights with `block_size = 32` are supported; projection biases, bfloat16, and non-default `accuracy_level` values are not implemented. |
| |
| See the [ONNX Runtime `MatMulNBitsQkv` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.MatMulNBitsQkv) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `A` | `aT` | `T1` | — | — | Shared activation of rank 2 `(M, K)` or rank 3 `(batch, sequence, K)`; only the last axis is the reduction axis. | required | |
| | `skip` | `skipT` | `T1` | — | — | Residual added to A before the normalization, with A's shape. | optional | |
| | `norm_scale` | `normScaleT` | `T1` | `1` | — | Simplified-layer-normalization (RMS) gain of shape `[K]`. | required | |
| | `q_B` | `qBT` | `T2` | `3` | — | Bit-packed uint8 Q weights of shape `(Nq, k_blocks, blob_size)`. | required | |
| | `q_scales` | `qScalesT` | `T1` | `2` | — | Per-block Q scales of shape `(Nq, k_blocks)`. Quantization is symmetric: there is no zero-point input, so codes are offset by the midpoint `2^(bits - 1)`. | required | |
| | `k_B` | `kBT` | `T2` | `3` | — | Bit-packed K weights of shape `(Nkv, k_blocks, blob_size)`. | required | |
| | `k_scales` | `kScalesT` | `T1` | `2` | — | Per-block K scales of shape `(Nkv, k_blocks)`. | required | |
| | `v_B` | `vBT` | `T2` | `3` | — | Bit-packed V weights of shape `(Nkv, k_blocks, blob_size)`. | required | |
| | `v_scales` | `vScalesT` | `T1` | `2` | — | Per-block V scales of shape `(Nkv, k_blocks)`. | required | |
|
|
| ## Outputs |
|
|
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `Q` | `qT` | `T1` | same as `A` | derived; see description | Query projection: A's leading axes with a trailing Nq. | required | |
| | `K` | `kT` | `T1` | same as `A` | derived; see description | Key projection: A's leading axes with a trailing Nkv. | required | |
| | `V` | `vT` | `T1` | same as `A` | derived; see description | Value projection: A's leading axes with a trailing Nkv. | required | |
| | `input_skip_bias_sum` | `residualT` | `T1` | same as `A` | same as `A` | The residual sum A + skip, with A's shape. Requires the skip input. | optional | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `accuracy_level` | `0` | Minimum internal accuracy level, following MatMulNBits semantics; this implementation supports the standard default 0. | |
| | `bits` | `4` | Bit width used to quantize all three weight matrices; only 4 is supported. | |
| | `epsilon` | `9.999999974752427e-7` | Epsilon used by the simplified layer-normalization reduction. | |
| | `K` | — | Input feature dimension shared by the normalized input and all projection weights. | |
| | `Nq` | — | Output feature dimension of the Q projection. | |
| | `Nkv` | — | Output feature dimension shared by the K and V projections. | |
| | `block_size` | — | Size of each quantization block along K; only 32 is supported. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T1` | `float32`, `float16` | |
| | `T2` | `uint8` | |
|
|
| ## 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 |
| - [`matmul-nbits-fused-rms-norm.wgsl.jinja`](build/webgpu/matmul-nbits-fused-rms-norm.wgsl.jinja) |
| - [`qkv-projection.wgsl.jinja`](build/webgpu/qkv-projection.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.MatMulNBitsQkv", { version: 1 }); |
| const { qT, kT, vT } = await kernel({ |
| aT: { data: aTData, shape: [2, 32] }, |
| normScaleT: { data: normScaleTData, shape: [32] }, |
| qBT: { data: qBTData, shape: [5, 1, 16] }, |
| qScalesT: { data: qScalesTData, shape: [5, 1] }, |
| kBT: { data: kBTData, shape: [3, 1, 16] }, |
| kScalesT: { data: kScalesTData, shape: [3, 1] }, |
| vBT: { data: vBTData, shape: [3, 1, 16] }, |
| vScalesT: { data: vScalesTData, shape: [3, 1] }, |
| }, { |
| attrs: { |
| K: 32, |
| Nq: 5, |
| Nkv: 3, |
| block_size: 32, |
| }, |
| }); |
| ``` |
|
|
