feat(train): add auto-scaling option for Multi-GPU Training

[nijkah]

What this does

This PR adds an opt-in auto-scaling feature for distributed training with Accelerate. When enabled via --auto_scale=true and running with multiple processes (GPUs), LeRobot:

• Multiplies the learning rate by world size (linear LR scaling).
• Divides the number of training steps by world size (ceil-div).

This keeps total sample count roughly comparable and preserves training dynamics while improving throughput.

How it was tested

Examples:

accelerate launch \
  --multi_gpu \
  --num_processes=2 \
  $(which lerobot-train) \
  --dataset.repo_id=${HF_USER}/my_dataset \
  --policy.type=act \
  --policy.repo_id=${HF_USER}/my_trained_policy \
  --output_dir=outputs/train/act_multi_gpu \
  --job_name=act_multi_gpu \
  --wandb.enable=true \
  --auto_scale=true

Tested

• A6000 x 2 GPU training

[Copilot]

Pull Request Overview

This PR introduces an auto-scaling feature for multi-GPU training that automatically adjusts learning rates and training steps when using multiple processes. When enabled via --auto_scale=true, the system multiplies the learning rate by the number of GPUs and divides training steps proportionally to maintain consistent total sample processing and training dynamics.

Key changes:

• Added auto_scale boolean flag to training configuration with logic to scale LR and steps based on world size
• Created comprehensive test suite to verify auto-scaling behavior with multi-GPU setups
• Updated documentation to explain the new auto-scaling feature and its usage

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
src/lerobot/configs/train.py	Adds auto_scale configuration field with documentation
src/lerobot/scripts/lerobot_train.py	Implements auto-scaling logic for learning rates and training steps
tests/training/test_auto_scale.py	Adds test case for auto-scaling with 2 GPUs
docs/source/multi_gpu_training.mdx	Updates documentation to describe auto-scaling feature and usage

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[pkooij]

Hi thank you for this PR, very nice addition. I am very busy this week but I will take a look asap next week!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.