File size: 5,100 Bytes
3d17c9b
8d4a681
3d17c9b
8d4a681
 
 
 
3d17c9b
8d4a681
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
---
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 | 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 normalization, with `A`'s shape. Requires `norm_scale`. | optional |
| `norm_scale` | `normScaleT` | `T1` | `1` | — | Simplified-layer-normalization (RMS) gain of shape `[K]`. Absent means the projections read `A` unnormalized. | optional |
| `gate_B` | `gateBT` | `uint8` | `3` | — | Bit-packed uint8 gate weights of shape `(N, k_blocks, blob_size)`. | required |
| `gate_scales` | `gateScalesT` | `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 |
| `gate_bias` | `gateBiasT` | `T1` | `1` | — | Optional gate bias of shape `[N]`, added before the activation. | optional |
| `up_B` | `upBT` | `uint8` | `3` | — | Bit-packed up weights, same shape and packing as gate_B. | required |
| `up_scales` | `upScalesT` | `T1` | `2` | — | Per-block up scales of shape `(N, k_blocks)`. | required |
| `up_bias` | `upBiasT` | `T1` | `1` | — | Optional up bias of shape `[N]`, added before the product. | optional |

## Outputs

| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- |
| `Y` | `yT` | `T1` | same as `A` | derived; see description | Gated MLP output: A's leading axes with a trailing N. | 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: 0 (unset), 1 (float32), 2 (float16), 3 (bfloat16), or 4 (int8). |
| `bits` | `4` | Bit width used to quantize both weight matrices; this implementation supports 2, 4, and 8. |
| `epsilon` | `0.00001` | Epsilon used by the optional fused RMS normalization. |
| `K` | — | Input feature dimension shared by both quantized weight matrices. |
| `N` | — | Output feature dimension shared by both quantized weight matrices. |
| `activation` | — | Activation applied to the gate projection; this implementation supports `silu`. |
| `block_size` | — | Size of each quantization block along K. |

## Type constraints

| Variable | Allowed dtypes |
| --- | --- |
| `T1` | `float32`, `float16` |

## 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)
- [`mlp-gate-up.wgsl.jinja`](build/webgpu/mlp-gate-up.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.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",
  },
});
```