Installation¶

Start with the environment that matches your workflow, then add the task and sandbox dependencies you need.

Non-training workflows: a standard Python 3.10+ environment.
RL training: a compatible verl Docker image with the training dependencies.

Install Uni-Agent¶

Clone the repository and enter its directory:

git clone https://github.com/verl-project/uni-agent.git
cd uni-agent

For RL training, install the bundled verl source into your training environment:

git submodule update --init --recursive
pip install --no-deps -e ./verl

Optional Dependencies¶

The additional dependencies introduced by Uni-Agent mainly include task dependencies and sandbox dependencies. Most are lightweight and can be installed on demand.

Task Dependencies¶

Task dependencies provide task-specific datasets, verifiers, and reward implementations. For example, install the swebench package only when running a SWE-Bench task:

SWE-Bench

pip install swebench

Sandbox Dependencies¶

Install the client package for the sandbox backend you plan to use, for example:

ModalveFaaS

pip install modal

pip install volcengine-python-sdk swe-rex

Ray Runtime Environments¶

For distributed Ray workloads, you can use a Runtime Environment to inject the required task, sandbox, and verl dependencies into every worker node.

working_dir: ./
excludes: ["/.git/"]
pip:
  packages:
    - "volcengine-python-sdk"
    - "swe-rex"
    - "swebench"
env_vars:
  PYTHONPATH: "verl"
  # ......

Pass the file when submitting the Ray job:

ray job submit --runtime-env runtime_env.yaml -- python entrypoint.py

Next, you can launch a sandbox and run some code.