How to specify the GPU used by vllm

[xiaolizh1]

trl/trl/trainer/grpo_trainer.py

Line 392 in a92e00e

if vllm_device.split(":")[0] == "cuda" and int(vllm_device.split(":")[1]) >= torch.cuda.device_count():

I have an 8-GPUs server, of which only the last two GPUs are available, and I set CUDA_VISIBLE_DEVICE=6,7, the value of torch.cuda.device_count() is 2. I want to load vllm into GPU 6, and I set vllm_device=cuda:6, but this line of code keeps giving an ValueError. What should I do?