| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.MatMulNBitsMlp |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Fuses a gated MLP over two block-quantized projections that share one activation: `Y = silu(A_norm @ gate + gate_bias) * (A_norm @ up + up_bias)`, using the `MatMulNBits` weight packing with no zero-point input. `A_norm` is `A`, `SimplifiedLayerNormalization(A, norm_scale)`, or `SkipSimplifiedLayerNormalization(A, skip, norm_scale)`, whose residual sum may be returned as a second output. Only `silu` and the default `accuracy_level = 0` are implemented; bfloat16 is not implemented. |
| |
| See the [ONNX Runtime `MatMulNBitsMlp` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.MatMulNBitsMlp) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `aT` | `A` | `T1` | — | — | Shared activation of rank 2 `(M, K)` or rank 3 `(batch, sequence, K)`; only the last axis is the reduction axis. | required | |
| | `skipT` | `skip` | `T1` | — | — | Residual added to `A` before normalization, with `A`'s shape. Requires `norm_scale`. | optional | |
| | `normScaleT` | `norm_scale` | `T1` | `1` | — | Simplified-layer-normalization (RMS) gain of shape `[K]`. Absent means the projections read `A` unnormalized. | optional | |
| | `gateBT` | `gate_B` | `uint8` | `3` | — | Bit-packed uint8 gate weights of shape `(N, k_blocks, blob_size)`. Bound in the packed storage layout: four blob bytes per u32 word. | required | |
| | `gateScalesT` | `gate_scales` | `T1` | `2` | — | Per-block gate scales of shape `(N, k_blocks)`, with the same dtype as `A`. Quantization is symmetric: this operator has no zero-point input, so codes are offset by the midpoint `2^(bits - 1)`. | required | |
| | `gateBiasT` | `gate_bias` | `T1` | `1` | — | Optional gate bias of shape `[N]`, added before the activation. | optional | |
| | `upBT` | `up_B` | `uint8` | `3` | — | Bit-packed up weights, same shape and packing as gate_B. Bound in the packed storage layout: four blob bytes per u32 word. | required | |
| | `upScalesT` | `up_scales` | `T1` | `2` | — | Per-block up scales of shape `(N, k_blocks)`. | required | |
| | `upBiasT` | `up_bias` | `T1` | `1` | — | Optional up bias of shape `[N]`, added before the product. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `yT` | `Y` | `T1` | same as `aT` | derived | Gated MLP output: A's leading axes with a trailing N. | required | |
| | `residualT` | `input_skip_bias_sum` | `T1` | same as `aT` | same as `aT` | 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 | |
| | --- | --- | --- | |
| | `K` | — | Input feature dimension shared by both quantized weight matrices. | |
| | `N` | — | Output feature dimension shared by both quantized weight matrices. | |
| | `accuracy_level` | `0` | Minimum internal accuracy level: 0 (unset), 1 (float32), 2 (float16), 3 (bfloat16), or 4 (int8). | |
| | `activation` | — | Activation applied to the gate projection; this implementation supports `silu`. | |
| | `bits` | `4` | Bit width used to quantize both weight matrices; this implementation supports 2, 4, and 8. | |
| | `block_size` | — | Size of each quantization block along K. | |
| | `epsilon` | `0.00001` | Epsilon used by the optional fused RMS normalization. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T1` | `float32`, `float16` | |
|
|
| ## 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 |
| - [`matmul-nbits-fused-rms-norm.wgsl.jinja`](build/webgpu/matmul-nbits-fused-rms-norm.wgsl.jinja) |
| - [`mlp-gate-up.wgsl.jinja`](build/webgpu/mlp-gate-up.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.MatMulNBitsMlp", { version: 1 }); |
| const { yT } = await kernel({ |
| aT: { data: aTData, shape: [2, 16] }, |
| gateBT: { data: gateBTData, shape: [4, 2, 4] }, |
| gateScalesT: { data: gateScalesTData, shape: [4, 2] }, |
| upBT: { data: upBTData, shape: [4, 2, 4] }, |
| upScalesT: { data: upScalesTData, shape: [4, 2] }, |
| }, { |
| attrs: { |
| K: 16, |
| N: 4, |
| block_size: 8, |
| activation: "silu", |
| }, |
| }); |
| ``` |
|
|
