fine-tuning-serving-openpi

Copy the closest config in src/openpi/training/config.py, set your LeRobot repo_id and matching weight loader, then run compute_norm_stats.py before launching train.py with your config name. Always recompute norm stats after any dataset or transform change.

Run examples/convert_jax_model_to_pytorch.py with the JAX checkpoint directory, config name, and output path. First apply the required transformers patches from src/openpi/models_pytorch/transformers_replace into your installed transformers package.

Serving pi0.5 inference needs one GPU with about 24 GB VRAM, while JAX fine-tuning requires an A100 80GB using roughly 60 GB. PyTorch DDP training uses about 40 GB per GPU across 1-8 GPUs.

Every OpenPI config requires precomputed normalization statistics before training starts. Run uv run scripts/compute_norm_stats.py with your config name, and re-run it whenever the config, dataset, or transforms change.

OpenPI covers pi0, pi0-fast, and pi0.5 flow-matching VLA models with JAX and PyTorch backends for ALOHA, DROID, and LIBERO. Use OpenVLA-OFT instead when you need continuous action heads with LoRA or must reproduce OpenVLA-OFT paper results.

Exit code 137 indicates an out-of-memory kill. Set XLA_PYTHON_CLIENT_PREALLOCATE=false and XLA_PYTHON_CLIENT_ALLOCATOR=platform, or reduce XLA_PYTHON_CLIENT_MEM_FRACTION to limit JAX memory preallocation.