Skip to content

[webgpu] Apply Flash Attention if sliding window exceeds KV cache length #25594

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Aug 1, 2025
Merged

[webgpu] Apply Flash Attention if sliding window exceeds KV cache length #25594

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
7a82784
Apply flash attention if sliding window exceeds KV cache length
daijh Jul 30, 2025
ade51b0
Fix typo
daijh Jul 30, 2025
5f00414
Check sequence length
daijh Jul 30, 2025
52ce1a1
Resolve comments
daijh Jul 31, 2025
4153d11
Minor update comment
daijh Jul 31, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Resolve comments
  • Loading branch information
@daijh
daijh committed Jul 31, 2025
commit 52ce1a144d317ece1592bb2a30a7879102e9e21d
4 changes: 1 addition & 3 deletions onnxruntime/contrib_ops/webgpu/bert/flash_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -910,9 +910,7 @@ Status ComputeFlashAttentionDecodeVxReduce(onnxruntime::webgpu::ComputeContext&

Status ApplyFlashAttention(const Tensor* Q, const Tensor* K, const Tensor* V, const Tensor* attention_bias,
Tensor* output, const Tensor* past_key, Tensor* present_key, const Tensor* past_value, Tensor* present_value,
const WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context, int local_window_size) {
ORT_ENFORCE(local_window_size == -1, "Sliding window is not supported yet in FlashAttention.");

const WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context) {
ORT_RETURN_IF_ERROR(CopyKVCache(context, parameters, K, past_key, present_key, V, past_value, present_value));
const int present_sequence_length = parameters.is_gqa_ ? parameters.seqlen_present_kv_cache_ : parameters.total_sequence_length_;
if (parameters.sequence_length_ > 1) {
Expand Down
2 changes: 1 addition & 1 deletion onnxruntime/contrib_ops/webgpu/bert/flash_attention.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,7 +134,7 @@ class FlashAttentionDecodeVxReduceProgram final : public Program<FlashAttentionD

Status ApplyFlashAttention(const Tensor* Q, const Tensor* K, const Tensor* V, const Tensor* attention_bias,
Tensor* output, const Tensor* past_key, Tensor* present_key, const Tensor* past_value, Tensor* present_value,
const WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context, int local_window_size = -1);
const WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context);

bool CanApplyFlashAttention(const Tensor* bias, const Tensor* present_key, const Tensor* present_value,
const WebgpuAttentionParameters& parameters, onnxruntime::webgpu::ComputeContext& context);
Expand Down
4 changes: 3 additions & 1 deletion onnxruntime/contrib_ops/webgpu/bert/group_query_attention.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,7 +198,9 @@ Status GroupQueryAttention::ComputeInternal(onnxruntime::webgpu::ComputeContext&
Tensor* present_value = context.Output(2, present_kv_shape);
parameters.past_present_share_buffer_ = present_key != nullptr && present_value != nullptr && past_key != nullptr && past_value != nullptr && past_key->DataRaw() == present_key->DataRaw() && past_value->DataRaw() == present_value->DataRaw();

bool use_sliding_window = (local_window_size_ != -1 && local_window_size_ < parameters.seqlen_present_kv_cache_ && local_window_size_ < parameters.total_sequence_length_);
ORT_ENFORCE(parameters.total_sequence_length_ <= parameters.seqlen_present_kv_cache_, "Total sequence length cannot be greater than the existing KV cache length.");
// Use a sliding window if the input sequence exceeds the window's length.
bool use_sliding_window = (local_window_size_ != -1 && local_window_size_ < parameters.total_sequence_length_);
if (!do_rotary_ &&
head_sink == nullptr && !use_smooth_softmax_ &&
!use_sliding_window &&
Expand Down