Installation

RL-Kernel requires Python 3.10 or newer and PyTorch. CUDA builds require a working CUDA toolchain; ROCm builds require a compatible ROCm environment.

From Source

git clone https://github.com/RL-Align/RL-Kernel.git
cd RL-Kernel
pip install -e .

Optional Backends

pip install -e ".[cuda]"

pip install -e ".[rocm]"

pip install -e ".[vllm]"

Install the vLLM extra only on rollout or benchmark environments that need the vLLM runtime. Core CI and mocked integration tests do not require it.

Development Dependencies

pip install -e ".[dev]"
pip install -r requirements-docs.txt

Documentation Preview

mkdocs serve

Then open the local URL printed by MkDocs.