onnxruntime
[webgpu] support smooth softmax for non-FA GQA implementation
#25285

Merged

[webgpu] support smooth softmax for non-FA GQA implementation #25285

fs-eire merged 5 commits into main from fs-eire/webgpu-smooth-softmax

[webgpu] support smooth softmax for non-FA implementation

8d3f73ea

Add stub for smooth softmax in FlashAttention

55f364e2

fs-eire marked this pull request as draft 314 days ago

fs-eire changed the title ~~[webgpu] support smooth softmax for non-FA GQA implementation~~ [WIP][webgpu] support smooth softmax for non-FA GQA implementation 314 days ago

fs-eire force pushed from f85e572b to dd8d91a4 314 days ago

fs-eire force pushed from dd8d91a4 to a63fc649 313 days ago

fs-eire force pushed from a63fc649 to 2fed3b1e 313 days ago

fs-eire force pushed from 2fed3b1e to c201d496 313 days ago

fs-eire force pushed from c201d496 to 5845d0ed 312 days ago

fs-eire changed the title ~~[WIP][webgpu] support smooth softmax for non-FA GQA implementation~~ [webgpu] support smooth softmax for non-FA GQA implementation 312 days ago

fs-eire marked this pull request as ready for review 312 days ago

Add implementation of head sink and smooth softmax

3a7b54ff

fs-eire force pushed from 5845d0ed to 3a7b54ff 312 days ago

fs-eire requested a review from

copilot-pull-request-reviewer 312 days ago

copilot-pull-request-reviewer commented on 2025-07-05

guschmue requested changes on 2025-07-07

guschmue added ep:WebGPU

resolve comments

de25b3ea

Merge remote-tracking branch 'origin/main' into fs-eire/webgpu-smooth…

43a24e80

guschmue approved these changes on 2025-07-07

fs-eire merged 6d28e2d2 into main 310 days ago

fs-eire deleted the fs-eire/webgpu-smooth-softmax branch 310 days ago

Reviewers

guschmue

copilot-pull-request-reviewer

Assignees

No one assigned

Labels

ep:WebGPU

Milestone

No milestone

onnxruntime [webgpu] support smooth softmax for non-FA GQA implementation #25285 Merged

[webgpu] support smooth softmax for non-FA GQA implementation #25285

onnxruntime
[webgpu] support smooth softmax for non-FA GQA implementation
#25285

Merged