File size: 3,472 Bytes
bbbd4e6
cc4bd23
bbbd4e6
cc4bd23
 
 
 
bbbd4e6
cc4bd23
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
library_name: kernels
license: apache-2.0
tags:
- kernel
- webgpu
- wgsl
---
# com.microsoft.GemmaRotaryEmbedding

`com.microsoft`  ·  ONNX Runtime contrib operator  ·  contrib since_version 1

## Description

Fuses the Gemma rotary-embedding tail: computes `sin` and `cos` from float32 `emb`, casts them to float16, then evaluates `q * cos + q_rot * sin` and the corresponding expression for `k`. `emb` has shape `(batch, seq, dim)` and is broadcast over the head axis of the `(batch, heads, seq, dim)` operands. Each product is rounded to float16 before the addition.

See the [ONNX Runtime `GemmaRotaryEmbedding` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.GemmaRotaryEmbedding) for the reference semantics.

## Inputs

| Name | Bind key | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `emb` | `embT` | `U` | `float32` | `3` | — | Rotary angles with shape `(batch_size, seq_len, dim)`, shared by every head. | required |
| `q` | `qT` | `T` | same as logical dtype | `4` | — | Query state with shape `(batch_size, num_heads, seq_len, dim)`. | required |
| `q_rot` | `qRotT` | `T` | same as logical dtype | `4` | — | Half-rotated query state, same shape as `q`. | required |
| `k` | `kT` | `T` | same as logical dtype | `4` | — | Key state, same shape as `q`. | required |
| `k_rot` | `kRotT` | `T` | same as logical dtype | `4` | — | Half-rotated key state, same shape as `q`. | required |

## Outputs

| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- |
| `output1` | `output1T` | `T` | same as `q` | same as `q` | Rotary-embedded query, same shape as `q`. | required |
| `output2` | `output2T` | `T` | same as `q` | same as `q` | Rotary-embedded key, same shape as `q`. | required |

## Type constraints

| Variable | Allowed dtypes |
| --- | --- |
| `T` | `float16` |
| `U` | `float32` |

## Device requirements

Every implementation variant requires `shader-f16`; the package has no variant-level fallback without that capability.

## 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
- [`gemma-rotary-embedding.wgsl.jinja`](build/webgpu/gemma-rotary-embedding.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.GemmaRotaryEmbedding", { version: 1 });
const { output1T, output2T } = await kernel({
  embT: { data: embTData, shape: [1, 2, 4] },
  qT: { data: qTData, shape: [1, 1, 2, 4] },
  qRotT: { data: qRotTData, shape: [1, 1, 2, 4] },
  kT: { data: kTData, shape: [1, 1, 2, 4] },
  kRotT: { data: kRotTData, shape: [1, 1, 2, 4] },
});
```