saving checkpoints with config files

[gray kelp]

Didn’t want this to get lost in #help so putting it here, hope that’s okay! Basically the checkpoints don’t save with config files so i don’t know how to push them merged to HF

#1248698127373041707 message

[fluid imp]

oh yes we will need to investigate apologies on the delay

[gray kelp]

Sorry I meant to link my post in help though they are similar issues

#help message

Oh np that’s okay I was just panicking because I was trying to do a progress study on my project

[viscid zodiac]

+1 on this

[fluid imp]

Ok so it looks like we'll need more time on this, we'll give you an answer in 10 hours or so

[gray kelp]

Thanks! @viscid zodiac my extremely jank work around is to train the model with different max_steps and then when it saves that model which is kind of like a checkpointed model it saves like the full version. it is a different training run each time but its the best i got rn

[torn verge]

Wait I noticed you did not use FastLanguageModel.from_pretrained()

hence why there's no config file

checkpoints ye dont have configs

theyre technically not supposed to be used that way 🙂

[gray kelp]

Oh i see...when I tried to do FastLanguageModel.from_pretrained I think i also got an error that fastLanguagemodel didnt have a push_to_hub_merged method but I maybe used everything wrong

[torn verge]

oh wait