The requirement.txt might be wrong and it doesn't works for me

[LIGHTCHASER1]

flash-attn should be at version 2.5.8 cuzz I met the same issue like https://github.com/Dao-AILab/flash-attention/issues/966.
Additionally, you forget to give the version of vllm. I found that vllm==0.4.2 works.

A requirement.txt works post here and you can try it from scratch (conda create -n simpleRL python=3.10):

accelerate
bitsandbytes
datasets
deepspeed==0.15.0
einops
flash-attn==2.5.8
isort
jsonlines
loralib
optimum
packaging
peft
ray[default]==2.12.0
tensorboard
torch
torchmetrics
tqdm
transformers==4.46.1
transformers_stream_generator
wandb
wheel
word2number
vllm==0.4.2