Add a cuDNN SDPA decode tier to the ONNX standard Attention CUDA kernel (Phase 1) - #29715

Merged

Ti-Tai Wang (titaiwangms) merged 12 commits into

mainfrom

copilot/add-cudnn-sdpa-decode-tier

Jul 29, 2026

Commits on Jul 14, 2026

Commits on Jul 16, 2026

Add cuDNN SDPA decode-tier tests (M1-M4) + non-skippable canary

Copilot
and
Copilot
committed
f100dd2
Add multi-batch nonpad_kv_seqlen rejection test for ONNX Attention (CPU)

Copilot
and
Copilot
committed
34767c6
Add ort-transformers-gpu-pytest skill (GPU pytest workflow + SDPA dispatch verification)

Copilot
and
Copilot
committed
dc1ab7d

Commits on Jul 17, 2026

Fix CUDA Plugin EP build: use dual-build-safe GetOrtStream in RunCudnnSdpaAttention

Copilot
and
Copilot
committed
df87b1a

Commits on Jul 21, 2026

Address tianleiwu review: key cudnn_decode_supported() probe by actual head config

Copilot
and
Copilot
committed
76f8cad

Commits on Jul 27, 2026

Address tianleiwu review round 2: warmup rationale, CUDA-graph capture

Copilot
and
Copilot
committed
00a3e44

Commits on Jul 28, 2026

Merge branch 'main' into copilot/add-cudnn-sdpa-decode-tier
Copilot
committed
c85130d

Commits on Jul 29, 2026

Fuse fully-masked-batch zero-fill into output transpose (cuDNN decode 4D path)
Copilot
committed
a28ceb8