AWS published a walkthrough for running distributed training on Amazon EKS using Ray Train Deep Learning Containers (DLCs) with Elastic Fabric Adapter (EFA). The Ray Train DLC includes the complete CUDA-through-EFA stack, so users do not need to build their own image. The reproducible sample fine-tunes Qwen2.5-1.5B with PyTorch FSDP spread across eight GPUs on two EFA-enabled g6.12xlarge instances, using the container exactly as published.

Why it matters

By shipping a prebuilt container that covers the networking and GPU stack, the approach aims to reduce setup overhead for multi-node, multi-GPU fine-tuning workloads.

Who should care

Teams running distributed training or fine-tuning on AWS infrastructure, particularly those working with PyTorch FSDP, Ray, and EKS, may find the sample directly applicable.