[Dependency Conflict] vLLM version mismatch between setup.py and requirements_train.txt

[showshout]

I am attempting to train Qwen2.5-VL using the provided scripts (e.g., train_grpo_qwen2_5_vl_7b.sh). I have encountered a blocking issue caused by a dependency conflict between setup.py and requirements_train.txt, which leads to a runtime error with the model configuration.

requirements_train.txt specifies vllm==0.8.2.

setup.py restricts the version to vllm<=0.6.3
When running with vllm==0.6.3 (as enforced by setup.py during installation), the code fails to load the Qwen2.5-VL model. It seems the older vLLM version cannot parse the newer Hugging Face config structure.

Although requirements_train.txt suggests using vllm==0.8.2, I cannot simply pip install this version. Upgrading to vllm==0.8.2 forces an upgrade of PyTorch (and likely CUDA versions), which I suspect will break the compatibility with the main verl codebase or other pinned dependencies

Is there a known "safe" combination of vllm, torch, and verl that supports Qwen2.5-VL?
Since setup.py and requirements.txt have conflicting versions for vllm (and potentially other libs), knowing the exact combination of torch, vllm, transformers, and flash-attn that you used to successfully train Qwen2.5-VL would be incredibly helpful. Thanks!