---
library_name: kernels
license: apache-2.0
tags:
- kernel
- webgpu
- wgsl
---
# com.microsoft.SparseAttention

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

## Description

Block-sparse causal attention used by Phi-3-small. `block_row_indices` and `block_col_indices` encode one or more CSR block masks, and layouts cycle over query heads. Grouped-query heads, separate or packed `[Q|K|V]`, explicit scaling, partial or full rotary embedding in NeoX or interleaved layout, and float16 are supported. The past/present key and value tensors share allocations and are updated in place. Head sizes must be non-zero multiples of 8; bfloat16 is not implemented.

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

## Inputs

| Name | Bind key | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `query` | `queryT` | `T` | same as logical dtype | `3` | — | Query `(batch_size, sequence_length, num_heads * head_size)`, or packed `[Q\|K\|V]` `(batch_size, sequence_length, (num_heads + 2 * kv_num_heads) * head_size)` when `key` and `value` are omitted. | required |
| `key` | `keyT` | `T` | same as logical dtype | `3` | — | Key `(batch_size, sequence_length, kv_num_heads * head_size)`. Omitted for packed QKV. | optional |
| `value` | `valueT` | `T` | same as logical dtype | `3` | — | Value `(batch_size, sequence_length, kv_num_heads * head_size)`. Omitted for packed QKV. | optional |
| `past_key` | `pastKeyT` | `T` | same as logical dtype | `4` | — | Key cache `(batch_size, kv_num_heads, max_cache_sequence_length, head_size)`, updated in place. | required |
| `past_value` | `pastValueT` | `T` | same as logical dtype | `4` | — | Value cache with the same shape as `past_key`, updated in place. | required |
| `block_row_indices` | `blockRowIndicesT` | `M` | `int32` | `2` | — | CSR row pointers `(num_layout, max_blocks + 1)`. Each layout starts at zero, is monotonically non-decreasing, and ends no later than that layout's `block_col_indices` width. | required |
| `block_col_indices` | `blockColIndicesT` | `M` | `int32` | `2` | — | CSR column indices `(num_layout, max_nnz_blocks)`, right-padded past each layout's non-zero count. Every active entry is in `[0, max_blocks)`. | required |
| `total_sequence_length` | `totalSequenceLengthT` | `M` | `int32` | — | — | Scalar or one-element vector holding the maximum total key length. Equal to `sequence_length` exactly in the prompt case, which is how the past length is decided. The value fits the cache, the sparse layout's `max_blocks * sparse_block_size` capacity, and the rotary-cache row count when rotary is enabled. | required |
| `key_total_sequence_lengths` | `keyTotalSequenceLengthsT` | `M` | `int32` | `1` | — | Per-batch total key length excluding padding, shape `(batch_size)`. Each value is at most `total_sequence_length` and is at least 1 for a prompt or at least `sequence_length` otherwise. | required |
| `cos_cache` | `cosCacheT` | `T` | same as logical dtype | `2` | — | Rotary cosine cache `(max_rotary_sequence_length, rotary_dimension / 2)`, where the width is a multiple of 8 no larger than `head_size / 2`. Required with `sin_cache` when `do_rotary` is 1. | optional |
| `sin_cache` | `sinCacheT` | `T` | same as logical dtype | `2` | — | Rotary sine cache with the same shape as `cos_cache`; required with it when `do_rotary` is 1. | optional |

## Outputs

| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
| --- | --- | --- | --- | --- | --- | --- |
| `output` | `outputT` | `T` | `3` | derived; see description | Attention output `(batch_size, sequence_length, num_heads * head_size)`. | required |
| `past_key` | `pastKeyT` | `T` | `4` | same as `past_key` | The key cache tensor itself after the in-place append; ONNX names this output `present_key`. | required |
| `past_value` | `pastValueT` | `T` | `4` | same as `past_value` | The value cache tensor itself after the in-place append; ONNX names this output `present_value`. | required |

## Attributes

Attributes and default values (overridable per request):

| Attribute | Default | Description |
| --- | --- | --- |
| `do_rotary` | `0` | Set to 1 to apply rotary embedding to Q and to K before it enters the cache; every other value disables rotary embedding. |
| `rotary_interleaved` | `0` | Set to 1 to rotate adjacent pairs instead of using the NeoX half-split; every other value selects the NeoX layout. |
| `num_heads` | — | Number of query heads. |
| `kv_num_heads` | — | Number of key/value heads; must divide `num_heads`. |
| `sparse_block_size` | — | Tokens per sparse block; one of 16, 32, 64, 128. |
| `scale` | — | Scale applied to query-key products; omitted or zero uses `1 / sqrt(head_size)`. |

## Type constraints

| Variable | Allowed dtypes |
| --- | --- |
| `T` | `float32`, `float16` |
| `M` | `int32` |

## 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, 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
- [`sparse-attention-sgmat.wgsl.jinja`](build/webgpu/sparse-attention-sgmat.wgsl.jinja)
- [`sparse-attention.wgsl.jinja`](build/webgpu/sparse-attention.wgsl.jinja)
- [`sparse-kv-append.wgsl.jinja`](build/webgpu/sparse-kv-append.wgsl.jinja)
- [`sparse-q-rotary.wgsl.jinja`](build/webgpu/sparse-q-rotary.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.SparseAttention", { version: 1 });
const { pastKeyT, pastValueT, outputT } = await kernel({
  queryT: { data: queryTData, shape: [1, 32, 8] },
  keyT: { data: keyTData, shape: [1, 32, 8] },
  valueT: { data: valueTData, shape: [1, 32, 8] },
  pastKeyT: { data: pastKeyTData, shape: [1, 1, 32, 8] },
  pastValueT: { data: pastValueTData, shape: [1, 1, 32, 8] },
  blockRowIndicesT: { data: blockRowIndicesTData, shape: [1, 3] },
  blockColIndicesT: { data: blockColIndicesTData, shape: [1, 3] },
  totalSequenceLengthT: { data: totalSequenceLengthTData, shape: [1] },
  keyTotalSequenceLengthsT: { data: keyTotalSequenceLengthsTData, shape: [1] },
}, {
  attrs: { num_heads: 1, kv_num_heads: 1, sparse_block_size: 16 },
});
```