Quick Start¶
This quick start requires only one machine with one A100 GPU. It runs Agent Lightning v1.0 with the local controller and provides the shortest path from an installed repository to a real rollout-driven training job.
Before you start¶
Complete Installation, including the verl GPU stack.
AGL v1.0 itself is lightweight, but policy inference and GRPO updates still require the GPU stack used by
verland vLLM.
1. Prepare the example¶
Download the Calc-X dataset from Google Drive, then extract it and place these files under examples/calc_x/data/:
Activate the project environment and install the dependencies:
source .venv/bin/activate
uv pip install openai httpx sympy \
"autogen-agentchat" "autogen-ext[openai]" \
"mcp>=1.11.0,<2" mcp-server-calculator
2. Start one local run¶
From the repository root:
The launcher performs four operations:
- starts Ray and the
verl/vLLM model backend; - starts
agl-serveron port8181; - starts
agl-controller runner_type=local; - runs the Calc-X training entrypoint.
Service logs are written under /tmp/.
Once the task is running, you can view the training results in W&B.
When you want to stop the run, press Ctrl+C once and wait for the script to exit. Do not press Ctrl+C repeatedly, as the cleanup process takes some time to stop all resources and processes safely.
What's Next¶
- Read Basics to learn the core Agent Lightning >= v1.0 concepts.
- Read the complete Calc-X example, which also covers the Kubernetes controller mode.