kapynInfrastructure

Fault tolerant distributed training on Amazon EKS using NVRx

NVIDIA Resiliency Extension (NVRx) enables fault‑tolerant distributed training on Amazon EKS. The integration with PyTorch FSDP allows asynchronous checkpointing and in‑process restarts, recovering from GPU faults in seconds. Benchmarks on H100 GPUs across 2 to 8 nodes show over 99 % training efficiency and rapid recovery.

AWS ML Blog·Sep 16, 2026

Opening Kapyn…