{
  "id": 16944164,
  "title": "Distributed training with Ray Train DLCs and EFA on EKS",
  "url": "https://urgent.news/2026/10/28/distributed-training-with-ray-train-dlcs-and-efa-on-eks",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-28T18:19:54.000Z",
  "source": {
    "name": "AWS Machine Learning",
    "slug": "aws-machine-learning",
    "url": "https://aws.amazon.com/blogs/machine-learning/distributed-training-with-ray-train-dlcs-and-efa-on-eks/"
  },
  "original_language": "en",
  "account": "The article discusses how the Ray Train Docker Container (DLC) streamlines distributed PyTorch training on a Kubernetes cluster.\n\nTraditionally, scaling out PyTorch training requires building a container with matching CUDA and PyTorch versions, NCCL library, EFA userspace libraries, AWS OFI NCCL plugin, and compatible Ray version. This is a complex task that doesn't relate to the model being trained. AWS simplifies this by providing the Ray Train DLC, which is a pre-built container with Ray's training libraries and the required dependencies.\n\nThe DLC image contains the distributed runtime with the Ray dashboard, job server, and the necessary training libraries. When the Ray Train DLC is used, it replaces the need to build and maintain the container. The container handles the networking, CUDA, NCCL, and other dependencies automatically.\n\nKubeRay, an open-source Kubernetes operator, simplifies the deployment of Ray clusters on Kubernetes. It creates the head and worker pods, starts Ray in the appropriate role for each, and manages the failure of pods. The Ray head coordinates the global control store, while the Ray worker nodes supply the GPUs and run the training processes.\n\nThe sample provided in the article demonstrates fine-tuning Qwen2.5-1.5B with PyTorch's Fully Sharded Data Parallel (FSDP) across eight NVIDIA L4 GPUs on two g6.12xlarge instances. The entire solution is contained within a single RayCluster manifest and environment variables file, with no need to build images. The Ray Train DLC allows training to run as a single cluster, leveraging the EFA stack for inter-node communication.",
  "summary": "The Ray Train Deep Learning Container ships the full CUDA-through-EFA stack, so you build no image. This post walks through a reproducible sample that fine-tunes Qwen2.5-1.5B with PyTorch FSDP across eight GPUs on two EFA-enabled g6.12xlarge instances on Amazon EKS, using the container exactly as published.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}