Hugging Face Accelerate — распределённое обучение PyTorch

★ 7.7 · devops

accelerate is a Claude Code skill that makes distributed PyTorch training accessible with just 4 lines of code, letting the same script run on a single GPU, multi-GPU nodes, or multi-machine clusters without rewriting core logic. It provides a unified API over DeepSpeed (ZeRO-2/3), FSDP, Megatron, and DDP, with automatic device placement and mixed-precision support for FP16, BF16, and FP8. Configuration is handled interactively via `accelerate config`, and any setup launches through a single `accelerate launch` command. The skill also covers gradient accumulation and integrates with the broader HuggingFace ecosystem (transformers, torch). It's aimed at ML engineers who need to scale training pipelines quickly without backend-specific rewrites.