Installation

Plato uses uv as its package manager, which is a modern, fast Python package manager that provides significant performance improvements over conda environments. To install uv, refer to its official documentation, or simply run the following commands:


curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env

To upgrade uv, run the command:


uv self update

To start working with Plato, first clone its git repository:


git clone git@github.com:TL-System/plato.git
cd plato

You can run Plato using uv run, using one of its configuration files:


uv run plato.py -c configs/MNIST/fedavg_lenet5.toml

In order to run any of the examples, first run the following command to include all global Python packages in a local Python environment:


uv sync

In case you need optional dependency groups, you can install them with:


uv sync --all-extras

or:


uv sync --extra mlx

where mlx is the name of the dependency group.

Useful extras in the current root package include:

llm_eval for server-side Lighteval evaluation
nanochat for Nanochat training and CORE evaluation
mlx for Apple Silicon MLX workloads
dp, rl, and mpc for specialized research workloads

Each example should be run in its own directory:


cd examples/server_aggregation/fedatt
uv run fedatt.py -c fedatt_FashionMNIST_lenet5.toml

This will make sure that any additional Python packages, specified in the local pyproject.toml configuration, will be installed first.

Optional: MLX Backend for Apple Silicon

To use MLX as a backend alternative to PyTorch on Apple Silicon devices, install the MLX dependencies:


uv sync --extra mlx

See the Quick Start guide for configuration details.

Optional: Server-side LLM Evaluation with Lighteval

To enable evaluation.type = "lighteval", install the evaluator stack:


uv sync --extra llm_eval

This installs lighteval together with the runtime dependencies used by Plato's built-in Lighteval adapter.

See:

Evaluation for the configuration contract
Server-side Lighteval for SmolLM2 for an end-to-end example

Optional: Nanochat Training and CORE Evaluation

To use Nanochat workloads or the nanochat_core evaluator, install:


uv sync --extra nanochat

Optional: SmolVLA + LeRobot Robotics Stack

The LeRobot / SmolVLA path is intentionally kept separate from the default Plato install so the root environment stays lean.

Migration note

Older runbooks may still reference uv sync --extra robotics. That root-package extra no longer exists. Use a dedicated environment that already has the LeRobot / SmolVLA stack installed, then verify it with:


uv run python -c "import lerobot; print(lerobot.__version__)"

See SmolVLA Trainer with LeRobot for the current setup guidance, configuration contract, and troubleshooting notes.

Building the `plato-learn` PyPi Package

The plato-learn PyPi package will be automatically built and published by a GitHub action workflow every time a release is created on GitHub. To build the package manually, follow these steps:

Clean previous builds (optional):
```
rm -rf dist/ build/ *.egg-info
```
Build the package:
```
uv build
```
Publish to PyPI:
```
uv publish
```
Or if you need to specify the PyPi token explicitly:
```
uv publish --token <your-pypi-token>
```

The uv tool will handle all the build process using the modern, PEP 517-compliant hatchling backend specified in pyproject.toml, making it much simpler than the old python setup.py sdist bdist_wheel approach.

Uninstalling Plato

Plato can be uninstalled by simply removing the local environment, residing within the top-level directory:


rm -rf .venv

Optionally, you may also clean uv’s cache:


uv cache clean

Optionally, you can also uninstall uv itself by following the official uv documentation.