Ray Train DLC and EFA run PyTorch FSDP fine-tuning across two EKS nodes
AIAWS Machine Learning Blog describes a sample that fine-tunes Qwen2.5-1.5B with PyTorch Fully Sharded Data Parallel across eight NVIDIA L4 GPUs on two g6.12xlarge instances. The job uses the Ray Train Deep Learning Container as published, with KubeRay running a single RayCluster manifest on Amazon EKS. Inter-node collectives run over Elastic Fabric Adapter (EFA), and no custom image is built.
Source details
AWS Machine Learning Blog · Sunita Koppar · aws.amazon.comPublished · added here
Original title: Distributed training with Ray Train DLCs and EFA on EKS