Ray Train — распределённое обучение моделей на кластере

★ 7.7 · orchestration

ray-train is a Claude Code skill that orchestrates distributed model training with Ray Train, scaling PyTorch, TensorFlow, and HuggingFace workloads from a single GPU to thousands of nodes with minimal code changes. It covers four core workflows: migrating existing PyTorch training loops to multi-GPU and multi-node clusters, fine-tuning HuggingFace Transformers models via TransformersTrainer across up to 16 workers, running hyperparameter sweeps with Ray Tune (up to 20 trials with ASHA early stopping), and implementing fault-tolerant checkpointing with automatic resume. Dependencies are ray[train], torch, and transformers. Ideal for ML engineers who need to train large models on a cluster or run distributed hyperparameter searches without rewriting their training logic.