fine-tuning-serving-openpi（微调服务 openpi）

OpenPI Fine-Tuning and Serving

End-to-end workflows for fine-tuning and serving Physical Intelligence's OpenPI models (pi0, pi0-fast, pi0.5) on robot manipulation tasks from the public openpi repository. Covers blank-machine setup, JAX training, PyTorch training, checkpoint conversion, and policy inference serving.

Quick start

Clone the public repo, install the workspace, then serve a pretrained policy:

git clone --recurse-submodules https://github.com/Physical-Intelligence/openpi.git
cd openpi
GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .
uv run scripts/serve_policy.py --env DROID

from openpi_client import websocket_client_policy

client = websocket_client_policy.WebsocketClientPolicy(host="localhost", port=8000)
result = client.infer(observation)
actions = result["actions"]  # numpy array of shape (chunk_size, action_dim)

Core concepts

Model family: OpenPI implements three model variants from Physical Intelligence:

Key design choices:

• Dual backend: JAX (primary, official training) and PyTorch (community, deployment-friendly)
• Config-driven: All training/serving parameters defined in src/openpi/training/config.py
• Norm stats: Every config requires precomputed normalization statistics before training
• WebSocket serving: Policy servers expose a WebSocket API for low-latency inference

Training loop invariant: After every config or dataset change, always re-run this cycle:

1. Compute norm stats → 2. Train → 3. Serve checkpoint → 4. Validate inference

Compute requirements

Workflow 0: Blank-machine setup

Copy this checklist and track progress:

Setup Progress:
- [ ] Step 1: Clone the public openpi repo with submodules
- [ ] Step 2: Install uv and sync the workspace
- [ ] Step 3: Install the editable package
- [ ] Step 4: Verify core imports and serving entrypoint

Step 1: Clone repo

git clone --recurse-submodules https://github.com/Physical-Intelligence/openpi.git
cd openpi

If you already cloned without submodules:

git submodule update --init --recursive

Step 2: Sync dependencies

GIT_LFS_SKIP_SMUDGE=1 uv sync

Step 3: Install editable package

GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

Step 4: Verify installation

uv run python -c "from openpi.training import config as _config; print(_config.get_config('pi05_droid').name)"
uv run scripts/serve_policy.py --help

When to use vs alternatives

Use this skill when:

• Fine-tuning pi0, pi0-fast, or pi0.5 on LeRobot or RLDS datasets
• Serving OpenPI policies for ALOHA, DROID, or LIBERO evaluation
• Converting JAX checkpoints to PyTorch format
• Debugging OpenPI training issues (norm stats, memory, config)

Use fine-tuning-openvla-oft instead when:

• Fine-tuning OpenVLA with continuous action heads and LoRA
• Reproducing OpenVLA-OFT paper results on LIBERO or ALOHA

Use evaluating-cosmos-policy instead when:

• Evaluating NVIDIA Cosmos Policy on simulation benchmarks

Workflow 1: JAX fine-tuning on LeRobot data

Copy this checklist and track progress:

JAX Fine-Tuning Progress:
- [ ] Step 1: Select and copy closest training config
- [ ] Step 2: Update dataset mapping and base checkpoint
- [ ] Step 3: Compute normalization statistics
- [ ] Step 4: Launch JAX training
- [ ] Step 5: Serve checkpoint and run inference sanity check

Step 1: Select config

Copy the closest config from src/openpi/training/config.py:

Step 2: Update dataset and transforms

# In src/openpi/training/config.py, modify your config:
TrainConfig(
    name="my_custom_config",
    model_type="pi05",
    data=LeRobotDataConfig(
        repo_id="your-org/your-dataset",
        # Adjust transforms to match your data format
    ),
    weight_loader=Pi05WeightLoader(),  # Match model type
)

Set repo_id for your dataset and ensure weight_loader matches the model type (pi0 vs pi0.5).

Step 3: Compute normalization statistics

uv run scripts/compute_norm_stats.py --config-name <config_name>

This must run before every training launch when config, dataset, or transforms change.

Step 4: Launch JAX training

XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 uv run scripts/train.py <config_name> \
  --exp-name=<run_name> \
  --overwrite

For full DROID RLDS training, add the rlds dependency group:

uv run --group rlds scripts/compute_norm_stats.py \
  --config-name pi05_full_droid_finetune \
  --max-frames 10000000

XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 uv run --group rlds scripts/train.py \
  pi05_full_droid_finetune \
  --exp-name=<run_name> --overwrite

Step 5: Serve and validate

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=<config_name> \
  --policy.dir=checkpoints/<config_name>/<run_name>/<step>

Verify with a test client:

from openpi_client import websocket_client_policy

client = websocket_client_policy.WebsocketClientPolicy(host="localhost", port=8000)
# Build observation matching your config's expected keys
obs = {"image": img_array, "state": state_array, "prompt": "pick up the cup"}
result = client.infer(obs)
print(f"Action shape: {result['actions'].shape}")  # (chunk_size, action_dim)

Workflow 2: PyTorch training and checkpoint conversion

Copy this checklist and track progress:

PyTorch Setup Progress:
- [ ] Step 1: Sync dependencies and verify transformer version
- [ ] Step 2: Apply OpenPI transformer patches
- [ ] Step 3: Convert JAX checkpoint to PyTorch format
- [ ] Step 4: Launch PyTorch training or serve converted checkpoint

Step 1: Sync dependencies

uv sync
uv pip show transformers

Step 2: Apply required patches

OpenPI PyTorch requires custom modifications to the installed transformers package:

cp -r ./src/openpi/models_pytorch/transformers_replace/* \
  .venv/lib/python3.11/site-packages/transformers/

Step 3: Convert JAX checkpoint

uv run examples/convert_jax_model_to_pytorch.py \
  --checkpoint_dir <jax_checkpoint_dir> \
  --config_name <config_name> \
  --output_path <pytorch_checkpoint_dir>

Step 4: Train or serve

Single GPU training:

uv run scripts/train_pytorch.py <config_name> --exp_name <run_name>

Multi-GPU distributed training:

uv run torchrun --standalone --nnodes=1 --nproc_per_node=<num_gpus> \
  scripts/train_pytorch.py <config_name> --exp_name <run_name>

Programmatic inference with converted checkpoint:

from openpi.training import config as _config
from openpi.policies import policy_config

config = _config.get_config("pi05_droid")
policy = policy_config.create_trained_policy(config, "<pytorch_checkpoint_dir>")
result = policy.infer(example)
actions = result["actions"]  # numpy array

Checkpoints follow the convention: checkpoints/<config_name>/<exp_name>/<step>/.

Workflow 3: Policy inference serving

Copy this checklist and track progress:

Inference Server Progress:
- [ ] Step 1: Choose target environment and checkpoint
- [ ] Step 2: Start policy server
- [ ] Step 3: Confirm server is reachable
- [ ] Step 4: Integrate client into robot or simulation code

Step 1: Choose environment

Default environment presets:

Step 2: Start server

Default mode (uses preset checkpoint):

uv run scripts/serve_policy.py --env ALOHA

Explicit checkpoint mode (custom or local model):

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_libero \
  --policy.dir=checkpoints/pi05_libero/my_run/20000

Add --default_prompt "task description" when runtime observations omit a prompt.

Step 3: Verify connectivity

uv run examples/simple_client/main.py --env DROID

Step 4: Embed remote client in robot code

Install the lightweight client in your robot environment:

pip install "openpi-client @ git+https://github.com/Physical-Intelligence/openpi.git#subdirectory=packages/openpi-client"

Full integration example:

from openpi_client import websocket_client_policy
import numpy as np

# Connect to remote policy server
client = websocket_client_policy.WebsocketClientPolicy(
    host="gpu-server.local", port=8000
)

# Build observation (keys must match policy transforms)
observation = {
    "image": np.random.rand(224, 224, 3),  # RGB image
    "state": np.zeros(7),                   # Joint positions
    "prompt": "pick up the red block",
}

# Get actions
result = client.infer(observation)
actions = result["actions"]  # shape: (action_chunk_size, action_dim)

# Execute first action on robot
robot.step(actions[0])

Common issues

Issue: Missing norm stats error

Fix: run scripts/compute_norm_stats.py --config-name <config_name> before training.

Issue: Out of memory during JAX training

Fix: set XLA_PYTHON_CLIENT_MEM_FRACTION=0.9, lower batch size, or configure fsdp_devices:

# In config: use model-parallel sharding
TrainConfig(
    ...
    fsdp_devices=4,  # Shard across 4 GPUs
)

Issue: OOM while loading PyTorch checkpoints

Fix: export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

Issue: Config not found

Fix: ensure config name exists in src/openpi/training/config.py (exact match from _CONFIGS dict).

Issue: PyTorch training diverges after library changes

Fix: reapply the transformer patch. Run uv cache clean transformers to reset, then reapply.

Issue: serve_policy.py crashes with ModuleNotFoundError

Fix: resync the public workspace first:

GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

If the missing module is simulator-related, install the extra runtime dependencies called for by that example:

uv pip install pytest robosuite==1.4.0 gym bddl easydict matplotlib

Issue: uv sync fails with rerun-sdk wheel mismatch

Fix:

uv sync --no-dev
# or
uv sync --no-dev --no-install-package rerun-sdk

Issue: Checkpoint download times out

Fix: install gsutil and prefetch manually:

pip install gsutil
gsutil -m cp -r gs://openpi-assets/checkpoints/pi05_libero /local/cache/

Remove stale .lock files if a previous download was interrupted.

Issue: Policy server exits with code 137

Fix: OOM kill. Set JAX memory variables:

export XLA_PYTHON_CLIENT_PREALLOCATE=false
export XLA_PYTHON_CLIENT_ALLOCATOR=platform

For HPC/cluster users

On Slurm-managed clusters, wrap commands with resource allocation:

srun --partition=gpu --gpus-per-node=1 --mem=64G --cpus-per-task=8 --pty bash

Route caches to scratch to avoid filling /home:

export HF_HOME=/scratch/$USER/.cache/huggingface
export XDG_CACHE_HOME=/scratch/$USER/.cache
export PIP_CACHE_DIR=/scratch/$USER/.cache/pip
export UV_CACHE_DIR=/scratch/$USER/.cache/uv

Avoid stacking cluster Python modules when using uv-managed environments. Typically module load cuda is sufficient.

Advanced topics

Config recipes and baselines: See references/config-recipes.md Training debugging guide: See references/training-debugging.md Checkpoint and environment mapping: See references/checkpoints-and-env-map.md Remote client integration: See references/remote-client-pattern.md PyTorch precision and patching gotchas: See references/pytorch-gotchas.md

Resources

• OpenPI repository: https://github.com/Physical-Intelligence/openpi
• OpenPI client package: https://github.com/Physical-Intelligence/openpi/tree/main/packages/openpi-client
• pi0 paper: https://www.physicalintelligence.company/blog/pi0
• LeRobot dataset format: https://huggingface.co/docs/lerobot