This example requires a single node with one GPU of at least 40GB memory.
- Download the data in parquet format from here and unzip it to the
datafolder:unzip calc-x-data.zip -d data. - Start ray:
bash ../../scripts/restart_ray.sh. To use Wandb, you need to set the WANDB_API_KEY environment variable before starting ray. - Run the agent:
python calc_agent.py. It automatically launches 4 agent workers by default. - In another terminal, launch the training server:
bash train.sh.
- The agent client will hang indefinitely if the environment is not properly configured. Check if uv and mcp are properly installed. Use
tests/test_mcp_calculator.pyto verify the installation.