Highlights
Denial of Service: 2 prior fixes. Scrutinize any change in this area.
applications/ColossalChat/coati/distributed/consumer.py: most-fixed (1 issue). Treat as high-risk during review.
0 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns
The bug types that recur here, drawn from past fixes, not open vulnerabilities.
Denial of Service: Mismatched FP16 gradient overflow states across distributed workers can cause training communication to hang indefinitely, resulting in a denial of service during synchronization.
Denial of Service: A GPU memory leak during model state synchronization can exhaust available VRAM, leading to out-of-memory crashes and service degradation.