onnxruntime
[WebGPU EP] Fuse QMoE 1-token decode path to reduce GPU dispatches
#27998
Merged
Merged
[WebGPU EP] Fuse QMoE 1-token decode path to reduce GPU dispatches #27998
guschmue
merged
12 commits into
main
from
opt/webgpu-qmoe-fused-1token
[WebGPU EP] Fuse QMoE 1-token decode path to reduce GPU dispatches
dea14482
fix incorrect results
54defba2
[WebGPU EP] Add per-row weight indirect support to DP4A SmallM and fi…
5978b03d
Merge branch 'main' into opt/webgpu-qmoe-fused-1token
ac84f9af
Remove ReplicateHidden: broadcast A row 0 via override_M
d70f7902
Remove per_row_weight_indirect: simplify to has_weight_idx_indirect
f90d9f15
Use num_tokens variable in Gate1Token dispatch for clarity
3ae0e4c2
Use M consistently for tile optimization checks instead of dispatch_M
4a953a70
Merge branch 'main' into opt/webgpu-qmoe-fused-1token
52d21598
Address PR review: separate dispatch_M uniform, skip zero-init for 1-…
c630eeff
Merge branch 'opt/webgpu-qmoe-fused-1token' of https://github.com/Mic…
96be6b51
copilot-pull-request-reviewer
commented
on 2026-04-10
Add WebGPU QMoE tests: EP block in RunQMoETest + 1-token path test
9d75e321
guschmue
approved these changes
on 2026-04-10
Reviewers
Assignees
No one assigned
Labels
ep:WebGPU
release:1.25.1
Milestone
No milestone
Login to write a write a comment.
Login via GitHub