Add large RAM GPU runner for vllm

[shermansiu]

vllm is a package that uses caching to enable high-throughput LLM inference. Building the CUDA wheel takes a lot of RAM which causes regular hosts to crash, which is why I'd like to request a bigger one for building the wheel.

Checklist:

• I want to request (or revoke) access to an opt-in CI resource:
  • Pinged the relevant feedstock team(s)
  • Added a small description explaining why access is needed

@conda-forge/vllm (I am one of the maintainers)

This can be merged after:

1. The maintainers (me and @maresb) get added as allowed open-gpu-server users:
   • shermansiu: Add shermansiu to the list of allowed users Quansight/open-gpu-server#63
   • maresb: Add @maresb to the list of allowed users Quansight/open-gpu-server#64
2. The vllm feedstock gets created: Add vllm staged-recipes#28931

[h-vetinari]

I'm fine to start off like this, but if vllm can be compiled with CPU-agents (compiling CUDA doesn't need a GPU), then we should do that.

There'll be one more PR to https://github.com/conda-forge/.cirun, which is also where we can then add further resource policies in the future

[h-vetinari]

conda-forge/.cirun#109

[shermansiu]

Thanks for creating the cirun PR! Do we need any additional PRs to use the larger CPU runners like ci_2xlarge?

[h-vetinari]

[...] https://github.com/conda-forge/.cirun, which is also where we can then add further resource policies in the future

☝️

Check out the pull request I linked. It creates a policy for vllm and adds it to a list of policies that are enabled for cirun-openstack-gpu-2xlarge runners. To be able to use other (e.g. smaller and/or CPU-only) runners, you need to add the vllm policy to the respective list.

[shermansiu]

Ah, so as I understand it, subsequent PRs will be made directly to https://github.com/conda-forge/.cirun and not https://github.com/conda-forge/admin-requests? Because now, the CI policy has been created?

[h-vetinari]

Yes. Please don't make me repeat myself N times