onnxruntime
webgpu: Enable FlashAttention for batched GQA with right-padded prompts
#29247

Merged

webgpu: Enable FlashAttention for batched GQA with right-padded prompts #29247

qjia7 merged 8 commits into microsoft:main from qjia7:feat/webgpu-flash-attention-batched

webgpu: fix GQA batched right-padded prefill with do_rotary

591df5b1

webgpu: support batch_size > 1 on the FlashAttention path

4c0492c7

Merge branch 'main' into feat/webgpu-flash-attention-batched

849ef988

webgpu: Tidy flash-attention seqlens_k handling and add prefill test

0c1512af

webgpu: Add non-flash attention path coverage to batched right-padded…

f624ecb4

webgpu: Simplify seqlens_k binding gate to use_seqlen_k_

9c5bff92

qjia7 requested a review from

copilot-pull-request-reviewer 15 days ago

copilot-pull-request-reviewer commented on 2026-06-25

webgpu: Use global total_sequence_length for indirect dispatch sizing

2adc29a7

qjia7 requested a review from

copilot-pull-request-reviewer 15 days ago

copilot-pull-request-reviewer commented on 2026-06-25

webgpu: Drop unused use_indirect_dispatch from DecodeVxReduce

9c23874f

qjia7 marked this pull request as ready for review 15 days ago

qjia7 requested a review from

guschmue 15 days ago

qjia7 requested a review from

hariharans29 15 days ago

guschmue approved these changes on 2026-06-25

hariharans29 approved these changes on 2026-06-26

qjia7 merged 92b4c663 into main 14 days ago

Reviewers

hariharans29

guschmue

Assignees

No one assigned

Labels

None yet

Milestone

No milestone