Projects

Training with Confidence: Catching Silent Errors in Deep Learning Training with Automated Proactive Checks

Based on OSDI 2025 paper by Yuxuan Jiang, Ziming Zhou, Boyu Xu, Beijie Liu, Runhui Xu, Peng Huang

My contributions after publication

• Designed and implemented the runtime checker with a watchdog to monitor trace file changes.
• Designed a new proxy to achieve torch.compile / distributed training compatibility and reduce runtime overhead.
• Revised the invariant inference and checking pipeline.