Skip to content

Changelog

Changelog for the Amazon Linux 2023-based Ray Train DLC images (train-ml-cuda).

Ray Train v1.0 — 2026-08-26

Tags: train-ml-cuda · train-ml-cuda-v1 · train-ml-cuda-v1.0 · train-ml-cuda-v1.0.0

Bundled versions: Ray 2.58.0 · PyTorch 2.13.0 · torchvision 0.28.0 · CUDA 13.0.2 · Python 3.13 · NCCL 2.29.7 · EFA 1.47.0 · GDRCopy 2.4.4 · flash-attn 2.8.3 · Transformer Engine 2.13.0 · DeepSpeed 0.19.2 · PyTorch Lightning 2.6.5 · Transformers 5.13.0

Highlights

  • Initial release of Ray Train DLC images on Amazon Linux 2023 — the PyTorch training stack plus Ray, replacing the upstream rayproject/ray-ml image deprecated at Ray 2.31.0
  • Ray 2.58.0 with the default, train, tune, and data extras; ray[default] supplies the dashboard and job-submission server used by ray job submit and KubeRay's health probes
  • PyTorch 2.13.0 with torchvision 0.28.0 on CUDA 13.0.2 and Python 3.13
  • EFA 1.47.0 with OpenMPI, the AWS NCCL OFI plugin, and GDRCopy 2.4.4 for multi-node collectives over EFA
  • flash-attn 2.8.3 and Transformer Engine 2.13.0 for fused attention and FP8 training; DeepSpeed 0.19.2 for ZeRO sharding
  • PyTorch Lightning 2.6.5 backing Ray Train's RayFSDPStrategy and RayDDPStrategy integrations, plus Transformers 5.13.0, datasets 5.0.0, accelerate 1.14.0, evaluate 0.4.6, and torchmetrics 1.9.0
  • One image serves both Ray head and Ray worker roles, so KubeRay's RayCluster spec drives the pods on EKS and HyperPod-EKS, and ray start drives them on EC2
  • Ports 6379 (GCS), 8265 (dashboard and job API), and 10001 (Ray Client) exposed
  • NCCL all_reduce_perf binary at /usr/local/bin/all_reduce_perf for verifying EFA connectivity
  • Pre-configured OpenSSH server (port 22) for inter-node MPI launches