Skip to content

Add CUDA Attention spec-coverage tests (GQA asymmetric head-size, output_qk, fp32 softcap+mask ordering) - #28371

Closed
titaiwangms wants to merge 5 commits into
microsoft:mainfrom
titaiwangms:fix-attention-spec-gaps-and-gqa-tests
Closed

Add CUDA Attention spec-coverage tests (GQA asymmetric head-size, output_qk, fp32 softcap+mask ordering)#28371
titaiwangms wants to merge 5 commits into
microsoft:mainfrom
titaiwangms:fix-attention-spec-gaps-and-gqa-tests

Fix RUFF formatting on output_qk assertion

449fa8c
Select commit
Loading
Failed to load commit list.
GitHub Advanced Security / CodeQL completed May 6, 2026 in 2s

1 configuration not found

Warning: Code scanning may not have found all the alerts introduced by this pull request, because 1 configuration present on refs/heads/main was not found:

API upload

  • ❓  <default>

New alerts in code changed by this pull request

  • 1 note

See annotations below for details.

View all branch alerts.

Annotations

Check notice on line 394 in onnxruntime/test/python/transformers/test_onnx_attention/common.py

See this annotation in the file changed.

Code scanning / CodeQL

Returning tuples with varying lengths Note test

attention_prompt_func returns
tuple of size 3
and
tuple of size 4
.