Installation

This guide will help you set up rLLM on your system.

Prerequisites

rLLM requires Python >= 3.10 (Python 3.11 is required if using the tinker backend).

Install uv (recommended)

Starting with v0.2.1, rLLM’s recommended dependency manager is uv. To install uv, run:

curl -LsSf https://astral.sh/uv/install.sh | sh

Install Python 3.11

Ensure that your system has a suitable installation of Python:

uv python install 3.11

Installation methods

Choose one of the following installation methods based on your needs.

Direct installation
Build from source
Docker
Without uv (not recommended)

Quick install with uv

Install rLLM directly from GitHub with a single command:

uv pip install "rllm[verl] @ git+https://github.com/rllm-org/rllm.git"

Replace verl with tinker to install with the tinker backend instead:

uv pip install "rllm[tinker] @ git+https://github.com/rllm-org/rllm.git"

Clone and install from source

For development or to access the latest features:

Clone the repository

git clone https://github.com/rllm-org/rllm.git
cd rllm

Create a virtual environment

uv venv --python 3.11
source .venv/bin/activate

Install rLLM with a training backend

rLLM supports two training backends: verl and tinker. Choose one based on your needs.

verl (GPU)
tinker (CPU/GPU)

Install with the verl backend for GPU-accelerated training:

uv pip install -e .[verl]

For CUDA 12.8 specifically:

uv pip install -e .[verl] --torch-backend=cu128

The verl extra installs vLLM by default. If you prefer SGLang for sampling rollouts:

uv pip install sglang --torch-backend=cu128

For AMD ROCm or Huawei Ascend accelerators, we strongly recommend installing rLLM on top of verl’s official Docker containers:

Install with the tinker backend for flexible training:

uv pip install -e .[tinker]

For CPU-only machines:

uv pip install -e .[tinker] --torch-backend=cpu

The tinker backend requires Python >= 3.11.

Activate your environment

Be sure to activate the virtual environment before running any scripts:

source .venv/bin/activate
python your_script.py

Containerized installation

For a containerized setup with GPU support:

Build the Docker image

docker build -t rllm .

Create and start the container

docker create --runtime=nvidia --gpus all --net=host \
  --shm-size="10g" --cap-add=SYS_ADMIN \
  -v .:/workspace/rllm -v /tmp:/tmp \
  --name rllm-container rllm sleep infinity

docker start rllm-container

Enter the container

docker exec -it rllm-container bash

The Docker setup includes GPU support via NVIDIA runtime and mounts the current directory to /workspace/rllm for easy development.

Install with pip and conda

While rLLM can be installed without uv, this is not recommended and may cause issues if you don’t have a compatible PyTorch or CUDA version preinstalled.

conda create -n rllm python=3.11
conda activate rllm
pip install -e .[verl]

This installation method may fail if you don’t have compatible CUDA and PyTorch versions already installed. Use uv for reliable dependency resolution.

Optional dependencies

rLLM provides additional optional dependencies for specific agent domains and framework integrations:

Framework integrations

sdk: LiteLLM proxy integration
smolagents: Hugging Face SmolAgents integration
strands: Strands agents framework

Domain-specific tools

web: Web agents (BrowserGym, Selenium, Firecrawl)
code-tools: Sandboxed code execution (E2B, Together)
swe: Software engineering tools (Docker, Kubernetes, SWEBench)
verifiers: Verifiers integration for validation

Development and utilities

dev: Development tools (pytest, ruff, mypy, mkdocs)
ui: UI components (httpx, python-multipart)
opentelemetry: OpenTelemetry SDK for observability

Install optional dependencies by adding them to the installation command:

# Install with web and code-tools extras
uv pip install -e .[verl,web,code-tools]

Advanced: Editable verl installation

If you wish to make changes to the verl backend itself:

git clone https://github.com/volcengine/verl.git
cd verl
git checkout v0.6.1
uv pip install -e .

Verify installation

Verify that rLLM is installed correctly:

python -c "import rllm; print(rllm.__version__)"

You should see the version number printed (e.g., 0.2.1).

Next steps

Quick start

Build your first math reasoning agent in 10 minutes

Core concepts

Learn about the key components of rLLM

Troubleshooting

If you encounter issues during installation:

Check Python version: Ensure you’re using Python >= 3.10 (3.11 for tinker)
GPU compatibility: Verify CUDA version matches PyTorch requirements
Dependency conflicts: Use uv instead of pip for better dependency resolution
GitHub issues: Check the GitHub issues page for known issues

For additional help, join our Slack community.

Get started

Tutorials

rLLM CLI & UI

Core concepts

Training backends

Guides

Unified workflow trainer

Advanced algorithms

Installation

Prerequisites

Install uv (recommended)

Install Python 3.11

Installation methods

Quick install with uv

Clone and install from source

Activate your environment

Containerized installation

Install with pip and conda

Optional dependencies

Advanced: Editable verl installation

Verify installation

Next steps

Quick start

Core concepts

Troubleshooting

Get started

Tutorials

rLLM CLI & UI

Core concepts

Training backends

Guides

Unified workflow trainer

Advanced algorithms

​Prerequisites

​Install uv (recommended)

​Install Python 3.11

​Installation methods

​Quick install with uv

​Clone and install from source

​Activate your environment

​Containerized installation

​Install with pip and conda

​Optional dependencies

​Advanced: Editable verl installation

​Verify installation

​Next steps

Quick start

Core concepts

​Troubleshooting

Prerequisites

Install uv (recommended)

Install Python 3.11

Installation methods

Quick install with uv

Clone and install from source

Activate your environment

Containerized installation

Install with pip and conda

Optional dependencies

Advanced: Editable verl installation

Verify installation

Next steps

Troubleshooting