onnxruntime
[WebGPU] Fuse Q/K RMSNorm into GroupQueryAttention for Qwen3-style models
#28484
Merged
Merged
[WebGPU] Fuse Q/K RMSNorm into GroupQueryAttention for Qwen3-style models #28484
hariharans29
merged
26 commits into
main
from
hari/webgpu_perf_2
[WebGPU] Fuse Q/K RMSNorm into GroupQueryAttention decode
d190ac19
[WebGPU] Extract RunLayerNormProgram helper
b9585c47
[WebGPU] Tighten GroupQueryAttentionPreNormFusion preconditions
68637297
[GQA] Reject q_norm_weight / k_norm_weight inputs on CPU/CUDA/JSEP
926437e7
Merge remote-tracking branch 'origin' into hari/webgpu_perf_2
98f40971
[Test] Use custom checker for SkipsAlreadyFusedNode
030d56c1
[Test] Add <cmath> include for std::abs
86d61e5b
[Test] Fix stale comment on BuildQwenQkPostNormPattern
c6965daa
[Optimizer] Update GroupQueryAttentionPreNormFusion class doc to refl…
b4bd290c
[Optimizer] Drop unused k{Q,K}NormWeightInputName constants
c263d6c5
[WebGPU] Refresh stale q/k_norm_weight comment in GQA
40986dfd
[WebGPU] Clarify qk_norm_epsilon_ comment covers prefill fallback too
6cbce92e
[Schema] Clarify q_norm_weight is honored by native WebGPU only
0fafb162
[Optimizer] Fix MatchPreNormReshapeChain doc comment (reshape order)
414684c1
[Test] Fix uniform_int_distribution assertion in GQA pre-norm fusion …
f247ae07
Address Copilot review comments on PR #28484
21e09aa5
Update docs
1ee43af1
Merge branch 'main' into hari/webgpu_perf_2
60debbaa
Merge remote-tracking branch 'origin/main' into hari/webgpu_perf_2
c69d3f44
Merge remote-tracking branch 'origin/main' into HEAD
e5e45a37
Merge main and resolve conflicts
6a3ed217
Left one out
0fe266e9
Copilot comments
2e0ef86c
Update doc
69312a3b
qjia7
dismissed these changes
on 2026-06-04
Add more tests
6a100f82
Fix
bf6e1860
qjia7
approved these changes
on 2026-06-09
Reviewers
Assignees
No one assigned
Labels
ep:WebGPU
Milestone
No milestone
Login to write a write a comment.
Login via GitHub