deepspeed universal checkpoint

1 article

Distributed Training Checkpoint Portability: Restore Matrix

A 2026 restore contract for distributed training checkpoint portability: state inventory, manifest, fault drills and a framework matrix covering PyTorch DCP, DeepSpeed UCP and Megatron-Core.

November 2, 2026• 31 min read
distributed training checkpoint portabilitypytorch distributed checkpointcheckpoint resharding