[14:32:14] <elukey>	 an example related to --^ (I had a chat with Dawid yesterday) - https://gerrit.wikimedia.org/r/c/operations/docker-images/production-images/+/1146891 is an example of base image that needs a GPU and a ton of memory to be built efficiently, and the ML team is structuring layers to avoid the daunting 4/5G compressed limit (nginx/swift). The build2xxx hosts are not equipped for this use case, the ml-lab10xx hosts are, so the proposal is to have 
[14:32:14] <elukey>	 docker-pkg run with push credentials on one of the ml-lab host, to build and push a repo like production-images only for ML to the registry.
[14:33:36] <elukey>	 this will unblock their primary use case that is having VLLM (https://github.com/vllm-project/vllm) usable as base image (like they do now with pytorch, but this one is more versed for LLMs)
[14:34:23] <elukey>	 and it disentangles the other issue with the registry (should they get a separate instance with different restrictions? Based on ceph/s3 etc..)
[14:34:44] <elukey>	 Cc: akosiaris, jayme (and whoever wants to chime in please do!)
[15:45:26] <akosiaris>	 I 'll have a closer look on Monday. From a high level overview, technical limitations aside, this sounds ok to me.