TierCheck: Tiered Checkpointing for Fault Tolerance in Large Language Model Training

In three linesTierCheck is a three-tier checkpointing system for LLM training. It maintains lightweight differential checkpoints in local/peer memory for fast recovery, asynchronously migrates base checkpoints to remote storage, and ensures global consistency without stalling training. On models up to 40B parameters, it reduces checkpointing time to under 10s.

Summary generated by Claude — human-verified