Skip to content

Installation

Claw-R1 uses veRL as its training backend and Ray for distributed execution.

Prerequisites

Requirement Version Notes
Python 3.10+ Conda recommended
CUDA 11.8+ Required for GPU training
Conda / Mamba latest For environment isolation
Git any For cloning repositories

Step 1 — Create the Environment

conda create -n clawr1 python=3.10 -y
conda activate clawr1

Step 2 — Install veRL

Claw-R1 requires the nightly version of veRL installed from source:

git clone https://github.com/volcengine/verl && cd verl
pip install --no-deps -e .
cd ..

Step 3 — Install Claw-R1

git clone https://github.com/AgentR1/Claw-R1 && cd Claw-R1
pip install -e .

Step 4 — Install Ray and FastAPI

pip install "ray[default]" fastapi uvicorn

Verify Installation

Run the following checks to ensure everything is installed correctly:

# Check veRL
python -c "import verl; print('veRL:', verl.__version__)"

# Check Ray
python -c "import ray; print('Ray:', ray.__version__)"

# Check Claw-R1 gateway
python -m claw_r1.gateway.gateway --help

GPU Memory

The Gateway Server is CPU-only and lightweight. Rollout workers (vLLM) and Training workers require separate GPU pools. See Async Training for multi-GPU configuration.

What's Next