onnxruntime
webgpu: Bump FlashAttentionDecodeQKV workgroup to 128, tile_size_k_vec to 32
#29586

Merged

webgpu: Bump FlashAttentionDecodeQKV workgroup to 128, tile_size_k_vec to 32 #29586

qjia7 merged 2 commits into main from webgpu-flash-attn-decode-qkv-wg128-tsk32

[WebGPU] FlashAttentionDecodeQKV: workgroup 128, tile_size_k_vec 32

64d58ee6

qjia7 marked this pull request as draft 40 days ago

[WebGPU] FlashAttentionDecodeQKV: gate 128/32 shape on m_tile==1

b4855b2f

qjia7 marked this pull request as ready for review 39 days ago

qjia7 requested a review from

hariharans29 39 days ago

qjia7 requested a review from

edgchen1 39 days ago

edgchen1 approved these changes on 2026-07-09

hariharans29 approved these changes on 2026-07-09

qjia7 merged 1ad2a1f2 into main 38 days ago

qjia7 deleted the webgpu-flash-attn-decode-qkv-wg128-tsk32 branch 38 days ago

Reviewers

hariharans29

edgchen1

Assignees

No one assigned

Labels

None yet

Milestone

No milestone