Expected way to set environment variables

[c-hoffmann]

I use vllm installed via pip on a linux server. And I want to set the vllm environment variables - if possible in a central file. open-webui for example uses a ".env"-file to set environment variables in a central file. Is there something similar for vllm? I have found the vllm/envs.py, but I'm not sure if I should set the variables there, since with an update it might get overwritten, right? Of course I can set up my own shell script with a lot of exports, but is that the intended way? How do you do it?

Thanks for your thoughts!

[c-hoffmann]

So I partially succeeded by creating ~/miniconda3/envs/vllm/etc/conda/actviate.d/env_vars.sh and ~/miniconda3/envs/vllm/etc/conda/deactviate.d/env_vars.sh. Inside the actviate.d/env_vars.sh I set a lot of environment variables, e.g.:

export CUDA_VISIBLE_DEVICES=1
export VLLM_LOGGING_LEVEL="WARNING"
export VLLM_LOGGING_PREFIX="WOOP"

and inside deactviate.d/env_vars.sh I deleted them again, e.g. with:

unset CUDA_VISIBLE_DEVICES
unset VLLM_LOGGING_LEVEL
unset VLLM_LOGGING_PREFIX

What works:

• CUDA_VISIBLE_DEVICES correctly sets my second GPU for usage with vLLM

What doesn't work:

• VLLM_LOGGING_LEVEL and VLLM_LOGGING_PREFIX do nothing. I expected to only see WARNINGs and worse (no INFOs) and have them logged with the prefix "WOOP", but no - I see all INFOs and no prefix.
  What am I doing wrong?

[c-hoffmann]

I found my mistake. Now the above solution works for all env variables. Still not sure if this is the optimal way. If you have a different approach, let me know.