Can I use two GPUs for VLLM?

[RL4LLM]

Thank you for your work.

I understand that you are using 1 GPU for VLLM inference and the other 7 GPUs for training.

Can I use two GPUs for VLLM inference? I noticed that using 1 GPU results in an out-of-memory (OOM) error when I set the max_completion_length to a large value.

[samma1570]

same problem

[tastelikefeet]

you can try SWIFT based on TRL and supports multi vllms and tensor-parallel