Fix Gemma4 CUDA EP: GQA head_dim guard, provider_options, opset lowering - #250

Closed

justinchuby wants to merge 10 commits into

mainfrom

fix-kv-shared-gqa

Commits on May 5, 2026

Guard Gemma4 GQA emission against head_dim > 256

justinchuby
and
Copilot
committed
62aa343
Fix Gemma4 CUDA EP: remove provider_options, use float masks

justinchuby
and
Copilot
committed
3dfb7e5
Lower opset 24→23 for non-default EPs to eliminate memcpy

justinchuby
and
Copilot
committed
6b77922
Update tests for empty CUDA provider_options behavior
justinchuby
committed
2a65db0
Remove GQA head_dim guard — new ORT supports head_dim=512
justinchuby
committed
4f4cbdf
Fix CUDA provider_options: always emit EP entry for GenAI
justinchuby
committed
de0f4a0
Apply suggestion from @justinchuby
justinchuby
authored
9d2edcb
Fix Gemma4 scale-free V norm FP16 overflow on CUDA

justinchuby
and
Copilot
committed
e50f7b6
Add stash_type=1 to RMSNormalization for FP32 variance accumulation

justinchuby
and
Copilot
committed
ea4e95c
Revert "Add stash_type=1 to RMSNormalization for FP32 variance accumulation"
justinchuby
committed
91acac7