Rafał Leszko
Rafał Leszko
Do I understand correctly that you'd like to have a setup that one runner can serve multiple request at a time? In other words, we would need to have a...
I think we scale differently for Realtime Video AI and for AI Jobs. - For Realtime Video, 1 runner always can handle 1 parallel stream - For AI Jobs, 1...
Ok, I'll fire a PR!
Here's a PR: https://github.com/livepeer/go-livepeer/pull/3639
Fine with me, but I see the CI build is failing.
> How is the `livepeerci/cuda:12.6.3-cudnn-devel-ubuntu22.04` created? I updated the Dockerfile.cuda-base but it does not seem to be creating it? This is the instruction: https://github.com/livepeer/go-livepeer/blob/280ea4dfe321143b83d51aafaa59adfdc27f0d76/docker/Dockerfile.cuda-base#L1 Do you have access to push...
I'll wait 1 week before merging this PR, because I noticed a few single cases with "insufficient balance". It was for high-resolution and the balance was at the borderline, so...
> This just streamlines the selection logic to avoid the for loop in `AISessionPool.Select` right? Hmm, I think it works differently for the Realtime Video AI and for AI Jobs....
> for try := 1; Ahh, I see it now. I think you're right. I removed it. Do you think it's ok now?
Holding off with merging this PR, because I still see `try=2` for our Realtime Video pipeline, which should not happen according tho the discussion above.