verl-e2e-testing

What problem does it solve?

It reduces the risk of breaking Megatron-Bridge model/provider integrations by verifying the provider end-to-end inside verl, including the external RL/training-loop wiring that Bridge-only tests may miss.

Core Features & Use Cases

• Runs a minimal verl Megatron PPO e2e: Executes a small Megatron backend job from a Bridge checkout to confirm the integration works end-to-end.
• Covers key compatibility surfaces: Validates provider configuration, HF import via Bridge, PEFT/LoRA wrapping, Megatron DDP wrapping, optimizer setup, and rollout/ref/critic wiring.
• Supports staged coverage: Moves from Level 0 smoke to optional save/resume, parallelism stress, optional Megatron-FSDP, architecture-specific checks, or a learning-signal run.

Quick Start

Start a Level 0 PPO e2e by exporting USE_MBRIDGE=True, VANILLA_MBRIDGE=False, VALUE_VANILLA_MBRIDGE=False, LORA_RANK=4, TOTAL_TRAIN_STEPS=1, and then running the verl wrapper script tests/special_e2e/run_ppo_trainer_megatron.sh with ENGINE=vllm.