Multi-GPU Training (HuggingFace)

Link: https://huggingface.co/docs/lerobot/multi_gpu_training

The guide referenced shows the user how to train policies on multiple GPUs using HuggingFace Accelerate

When you launch training with accelerate:

1. Lerobot detects if it’s running under accelerate
2. Your batch is automatically split across GPUs
3. Gradients are synchronized across GPUs during backpropagation
4. Main process logs to wandb and saves checkpoints