AWS adds Ray Train DLCs and EFA support for distributed training on EKS

AWS Ray Train DLC Adds EFA Support

AWS just dropped a Ray Train DLC that runs PyTorch across EKS nodes using EFA, cutting custom container builds. The image, tagged public.ecr.aws/deep-learning-containers/ray:train-ml-cuda-v1.1, bundles CUDA, NCCL, EFA libraries, and Ray training libraries in one unit. It auto-detects host drivers, flips on forward-compatibility if needed, and removes the need for separate head or worker images. KubeRay on EKS creates a head pod and workers that all use the same image, DLC defaults to FI PROVIDER=efa. A sample fine-tunes Qwen2.5-1.5B with PyTorch FSDP on eight NVIDIA L4 GPUs across two g6.12xlarge instances, all without rebuilding images. Because the DLC pre-installs the HuggingFace stack, jobs can install extra libraries on the fly, keeping the image unchanged. The cluster pulls the DLC via a NAT gateway, runs the head on an m7i.xlarge, and workers on g6.12xlarge GPUs. Training completes faster because EFA eliminates TCP overhead, letting developers focus on model code instead of container logistics. AWS's Ray Train DLC is now the go-to image for EKS-based PyTorch training, simplifying scaling and cutting operational friction.