The Workflow of Training with verl-omni main-diffusion Pipeline
This post is to review the overall workflow after launching a task from terminal bash examples/..../xxx.sh. We take FlowGRPO+SD3.5 Medium as an example here.
Table of Contents
1. Entrance
The entrance in the bash file is verl_omni/trainer/main_diffusion.py:main(). The main() parses configurations, validates them, and passes to run_diffusion(). The run_diffusion() resolves the configuration, initializes Ray backend, creates an instance of TaskRunner class, and puts this instance on Ray to automatically manage resources and run.
The core of TaskRunner class is its run() method, which models the complete training loop.