Support Attention(24)-CUDA and disjoint from contrib op#27542

Merged

titaiwangms merged 37 commits intomainfrom

titaiwang/design_attention_with_ai

Mar 9, 2026

Commits on Mar 4, 2026

ONNX Attention thin-dispatcher: direct flash/MEA/unfused dispatch with nonpad_kv_seqlen support

titaiwangms
and
Copilot
committed
Update Operator Kernel document
titaiwangms
committed
Fix cutlass FMHA crash when attention bias stride is unaligned

titaiwangms
and
Copilot
committed
Fix GQA decode eligibility, padding mask wiring, and 4D BNSH Q transpose

titaiwangms
and
Copilot
committed
Fix all-false mask crash in ConvertMaskToSeqlensKernel

titaiwangms
and
Copilot
committed
Replace host memcpy with device-side fill for CUDA graph capture

titaiwangms
and
Copilot
committed
Handle bool mask in decode path to support variable padding

titaiwangms
and
Copilot
committed
Fix NaN output for all-false bool masks in MEA path

titaiwangms
and
Copilot
committed
Fix cutlass FMHA bias alignment crash for unaligned kv_sequence_length

titaiwangms
and
Copilot
committed
Fix 3D mask test to use consistent per-batch padding semantics

titaiwangms
and
Copilot
committed
Remove 11 redundant GQA tests from test_gqa.py

titaiwangms
and
Copilot
committed
Fix padding mask bugs: zero present buffers, decode offset, MEA 2D expansion

titaiwangms
and
Copilot
committed
Add TODO for GQA unfused attention fallback

titaiwangms
and
Copilot
committed
Add TODO comments for GQA+float_mask and 4D present gaps

titaiwangms
and
Copilot
committed
Add TODO comments for softcap/softmax_precision and output_qk gaps

titaiwangms
and
Copilot
committed
Revert "Add TODO comments for GQA+float_mask and 4D present gaps"
titaiwangms
committed
Code cleanup: remove dead function, fix comments, CUDA-graph-safe 2D broadcast

titaiwangms
and
Copilot
committed
Add test improvements: unfused MHA, 4D BNSH GQA, broadcast mask, float mask tests

titaiwangms
and
Copilot
committed
lint
titaiwangms
committed

Commits on Mar 6, 2026

Commits on Mar 7, 2026

Commits on Mar 9, 2026

Merge branch 'main' into titaiwang/design_attention_with_ai
titaiwangms
committed