FlashInfer: diagnosing a CUDA Graph hang
A two-GPU reproducer isolated FTZ-sensitive sentinel handling; an upstream fix was backported.
Problem & contribution
H200 TP=2 rollout hung in a CUDA Graph path, surfacing as a hang or RPC timeout.
Built a two-GPU reproducer and compared eager/Graph and fused/unfused execution. Isolated FTZ-sensitive Lamport sentinel handling, backported the upstream bitwise fix and ran model-level Graph replay regression.
Validation
| Condition / metric | Result |
|---|---|
| Repeated Graph replay after the fix | No observed hang / timeout |
Completed root-cause isolation, backporting and regression validation.
Source: the author’s September 2026 resume and engineering summary; results apply to the stated test conditions.
This is an internship backport and validation, not authorship of the original upstream fix. The separately merged FlashInfer #5171 addresses installed-package tests and is recorded independently.
A separate upstream contribution
#5171 ↗ · Installed-package test compatibility, merged; distinct from the backport on this page.