Installation | NeMo Gym

Python 3.13.14 is required. Refer to Prerequisites for full system requirements.

Install from PyPI for the quickest setup. Clone from git if you need the latest features and environments, or use a NeMo RL container if you intend to train with NeMo Gym and NeMo RL.

PyPI

Git

Container

python3.13 -m pip install nemo-gym

Or with uv:

uv venv --python 3.13.14 && source .venv/bin/activate
uv pip install nemo-gym

The package includes built-in environments and CLI commands. Config and data paths resolve against the package install location automatically. To pin a release: python3.13 -m pip install nemo-gym==0.7.0.

Ray isn’t included by default. To run environments that use Ray, such as local_vllm_model or the SWE agents, install the ray extra: python3.13 -m pip install "nemo-gym[ray]".

Verify Installation

gym --version

You should see output like:

NeMo Gym v0.7.0
Python 3.13.14
Installation: /path/to/Gym

Next Steps

Quickstart

Run your first evaluation.

Component dependency locks

Repository-managed servers have their own pyproject.toml and committed uv.lock, enforced by CI. gym env start, gym env prefetch, and gym env test prefer uv sync --locked, including when uv_venv_dir places the virtual environment outside the checkout. Components are not combined into one workspace. Existing custom components with requirements.txt or an unlocked pyproject.toml retain their previous setup, discovery, and onboarding behavior without migration.

A stale component lock fails setup; sync errors do not trigger an unlocked retry. After intentionally editing managed dependencies, run uv lock in that directory and commit both files. Do not use --frozen to hide a stale lock.

If a successful locked sync does not satisfy inherited parent requirements, setup automatically uses compatibility resolution against the component’s dependency metadata without changing its committed project or lock. This propagates parent-only sandbox provider plugins (including their exact sources and dependencies) and the parent’s compatible OpenAI version. Ray constraints also act as overrides and apply only to components that require Ray. The resulting environment is checked again; unresolvable requirements or source mismatches still fail explicitly. Telemetry retains its selected extra. No user configuration changes are required.

Wheel installations carry the source metadata needed for relative Gym dependencies; the installer reconstructs that metadata beside the writable component environment without modifying site-packages or resolving a different Gym version from PyPI. Custom indexes and prerelease policy belong in the project used to generate the lock, not in a startup-only dependency override.

Benchmark preparation

A dataset that needs additional preparation dependencies declares prepare_project pointing at an isolated project with pyproject.toml and uv.lock. Gym syncs it with uv sync --locked and imports/runs the preparer in that environment, including its cache-freshness hook, without installing packages into the running Gym interpreter.

Existing dataset.prepare_dependencies configs remain supported unchanged. These arguments are passed to uv pip install --python <running Gym interpreter> before importing the preparer, including editable installs and requirements-file options. When nonempty, prepare_dependencies takes precedence over prepare_project, so existing config overrides keep their behavior. Use prepare_project for a reproducible locked environment; migrating existing preparation configs is optional.