You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
tests: flex_attention NoValidChoicesError in test_pd_stream_admission CUDA cases on torch 2.10 #294
With a GPU visible, three [cuda] cases in opt_prime/tests/test_pd_stream_admission.py fail on torch 2.10.0+cu128 with an inductor flex_attentionNoValidChoicesError. CI (CPU only) never runs them.
Evidence
Found on optimus-rm52ff463 on the 2x A100 development host (one A100 visible) while verifying #114 / PR #293. The failure is the same with and without TORCH_DEVICE_BACKEND_AUTOLOAD=0, so it is unrelated to the furiosa-torch autoload problem.
Scope
Find whether the flex_attention block mask / shapes used by the stream-admission path are outside what torch 2.10 inductor can compile on A100, and fix the call or fall back to eager / SDPA.
Done when
test_pd_stream_admission.py passes with a GPU visible on torch 2.10 (and 2.13), or the CUDA cases skip with a stated reason.
Activity
fix: run paged decode on torch 2.10 CUDA with pages under 64 tokens (#…