Distributed training with Ray Train DLCs and EFA on EKS

What happened

AWS demonstrates distributed PyTorch training on EKS using its published Ray Train DLC, KubeRay, and EFA without building a custom container. The sample runs Qwen2.5-1.5B with FSDP across eight NVIDIA L4 GPUs on two g6.12xlarge instances, using random token batches to verify training and cross-node communication. The guide covers cluster deployment, transport verification, workload configuration, and cleanup.

Read next

• More changes in today’s briefingGoogle DeepMind announces Gemini 4 Carbon with cybersecurity and long-horizon capabilities

The story so far

AWS demonstrated distributed PyTorch training on Amazon EKS using its published Ray Train Deep Learning Container, KubeRay, and Elastic Fabric Adapter (EFA), without building a custom container. The sample ran Qwen2.5-1.5B with Fully Sharded Data Parallel (FSDP) across eight NVIDIA L4 GPUs on two g6.12xlarge instances, using random token batches to verify training and communication between nodes. The guide covers cluster deployment, transport verification, workload configuration, and cleanup. Its demonstration verifies the training setup and cross-node communication using synthetic inputs.