[Feature Request] Don't save at every step

[ljleb]

Correct me if I'm wrong, but saving a checkpoint on disk at every step of adjusting merge ratios slows down the learning process by a lot IIUC. I think every n iterations should be treated as an epoch, i.e. it should be possible to save multiple models over time, but only save one intermediate model after a specified number of iterations.

Is it okay with you if I open a PR for this? Not sure how long I'll take as I'm unfamiliar with the repo.

[Xerxemi]

You're right about that, but actually, it saves the model to disk to be able to use it to generate but doesn't keep it, it's a limitation of the current code. Xynon is looking into swapping the testing backend for either runtime's or supermerger's but both of us are pretty busy rn (both aformentioned are different codebases so will take a while to integrate) and as such there is no eta.

Saving models as epochs wouldn't be useful, as all past final and test merges are saved to history tsvs already.

[ljleb]

Ah makes sense, as we know the original models we only need to save the receipe. simple and effective.