Multiple Critical Issues with Deepscaler repo

[troy12x]

Hi,

This repository is based on DeepScaler, which has several major issues that need urgent fixing:

Flash Attention 2.0 dtype mismatch:

"Error: Flash Attention 2.0 only supports torch.float16 and torch.bfloat16 dtypes, but the current dtype in Qwen2ForCausalLM is torch.float32.
Module Import Error (vllm._version missing):"

"Error: No module named 'vllm._version' from vllm.version import version as VLLM_VERSION
RuntimeError: Lack of CPU Swap Space:"

lead to
### ↓

Error: RuntimeError: Aborted due to the lack of CPU swap space
These issues are preventing proper execution, and they need to be fixed. Please address them as soon as possible.

[Pranjal2041]

We will update the verl version shortly after ensuring reproducibility. For now, can you try exporting VLLM_ATTENTION_BACKEND='XFORMERS'?

Reference