[PyTorch][torch.compile] Make get_attention_backend traceable without graph breaks#3189

Open

pggPL wants to merge 10 commits into

NVIDIA:mainfrom

pggPL:attention_get_attention_backend_traceable

Commits on Jul 8, 2026

Make get_attention_backend traceable by torch.compile without graph breaks
pggPL
committed
Return int from the fused-attn backend probe and test symbolic-int path
pggPL
committed
Replace the FusedAttnBackend dict with a python-side IntEnum
pggPL
committed
Soften the FusedAttnBackend enum transition following the DType pattern
pggPL
committed
Add FP8 backend-selection coverage to the torch.compile tests
pggPL
committed
Simplify the fused-attn probe wrapper and drop obsolete logging guards
pggPL
committed
Consolidate backend-selection compile tests; keep probe args literal
pggPL
committed
Drop references to unreleased dynamo features from comments
pggPL
committed
Assert selection results only in the compile test
pggPL
committed
Merge remote-tracking branch 'upstream/main' into attention_get_attention_backend_traceable
pggPL
committed