[13:06:48] hello team, as a followup to a slack discussion here are a couple of links to charts the we imported from upstream with minor changes. this is mostly a formality I would say as we were already using kserve for a very long time, these 2 charts simply add additional CRDs and 1 type of new LLM controller to kserve's capabilities. here they are: [13:06:48] https://wikitech.wikimedia.org/wiki/Helm/Upstream_Charts/kserve-llmisvc-crd-minimal [13:06:48] https://wikitech.wikimedia.org/wiki/Helm/Upstream_Charts/kserve-llmisvc-resources [13:06:48] we are currently testing them on ml-staging-codfw [13:13:08] dpogorzelski: thanks a lot for adding those! I have a couple of notes [13:14:19] Afaics from https://gerrit.wikimedia.org/r/c/operations/deployment-charts/+/1331389 the chart is very dependent on knative serving, I am not sure if we want to explicitly state that in Chart.yaml or not [13:15:18] the other thing is about https://gerrit.wikimedia.org/r/c/operations/deployment-charts/+/1331389/4/charts/kserve-llm-inference/README.md - the .wikimedia.org suffix is configurable, not hardcoded. From the README it seems something different, the chart seems to be tailored for one use case [13:15:26] that may be it, but better to clarify that point [13:15:55] and last but not least, please the next time ask somebody to review the code etc.. Self reviews are good but often the may be biased :) [13:16:55] sure thing i can fix both. to add on top, the knative dependence will be temporary as afterwards we will look into standalone deployments + gateway api to take advantage of application specific features like cache aware routing etc. https://phabricator.wikimedia.org/T436650 [13:17:50] okok nice [13:18:08] doing all of this at once felt a bit too much so i split this into 2 stages [13:18:54] totally agree yes [13:19:29] Anyway, if there is anybody willing to do a review of the above charts speak up! [13:19:43] just to verify common things, not a blocker for you [13:23:09] 👍