batch of openai fientuning jobs

[Butanium]

Hey, afaik the current repo doesn't support well launching e.g. 10 openai fientuning jobs on got 4.1 because you can only have 3 simultaneous runs at the same time. I managed this by having a custom loop that wait for a job to finish before submitting the next one, but I was wondering if you had such a module already implement somewhere else / if it'd make sense to have a pr for this.

Apparently /v1/fine_tuning/model_limits exists to get the number for the specific api key

[johny-b]

Nothing like this exists.

Generally these loops are pretty bad when you work in an org shared with others:

• You eat up all the training runs
• And also you are more likely to do useless training runs (at least that was true in my case, I used such loops several times and don't use them anymore)

OFC this could be great when you're the only person using an org.
So hm. I guess a PR could be fine for this? Do you have some specific interface in mind?

[Butanium]

hmmm that's fair, maybe default behavior should be use all available spots for the first batch, and then use MODEL_LIMIT - 1 simultaneous job so that your batch never blocks the org fully. You could also have a nice argument that if true only submit jobs if there are 2 free spots (kind of like a scavenger slurm queue)

You'd have an argument to override that and go full steam obviously