Back to Daytona

veRL ReTool Backend Benchmark

guides/python/reinforcement-learning/verl-retool/README.md

0.173.01.0 KB
Original Source

veRL ReTool Backend Benchmark

Overview

This directory contains the benchmark script used by Daytona's veRL guide. benchmark_tool_backends.py compares Daytona, Docker, and SandboxFusion backends from a local veRL checkout with the recipe submodule initialized. The Docker backend can also run standalone without a veRL checkout.

Requirements

  • A local veRL checkout with the recipe submodule initialized
  • A Python environment where veRL is already installed
  • Either DAYTONA_API_KEY or DAYTONA_JWT_TOKEN exported in your shell (for the Daytona backend)

Quick Start

From your veRL environment:

bash
cd /path/to/daytona/guides/python/reinforcement-learning/verl-retool
pip install -e .

Run the benchmark:

bash
python benchmark_tool_backends.py \
  --backend daytona \
  --verl-root /absolute/path/to/verl \
  --concurrency 1 4 8 16 32 64 128

The script runs simple_stdout, cpu_bound_stdout, and runtime_error, and writes summary.json and results.csv under outputs/daytona/<timestamp>/.