Setup a versioning system similar to revision from huggingface #2139
Description
I just saw that the latest fixes included a removal of the revision variable when using peft models.
If that fixes some bugs, okey.
But we need to think about atleast a similar system like that.
Just now when debugging my qwen tokenizer issue I first found that the qwen2VL 4bit quants where updated between me training an adapter and then doing inference.
While that was not the inherent issue for my problem it did make debugging a bit harder, as I was not able to test the older version of the model.
Even when giving the revision parameter, the newest version was downloaded (obv. now cause the revision parameter was removed)
I even tried to remove the version and upload an older version manually. Unsloth completly ignores that and downloads the newest version!
Not only should we as user have the ability to freeze the version of the base model we use, the library should warn us if the versions between adapter and base-model used are different. So we can atleast better debug, maybe even ensure not random differences in performance from one day to the next.