embeddings not preserved when resuming training after completition [Bug]:

[MartinoCesaratto]

What happened?

After training a LORA with multiple additional embeddings, I decided to do a few more epochs of training. Noticed that the samples were quite different, and after some testing I can say that quite certainly it retained the LORA but started training the embeddings from scratch. If I explicitly tell it to load embeddings from a file I can correctly load the latest ones, but it shoulden't actually load them from file since it already got the trained ones in the backup.

For context, the training started from readily-trained embeddings, but I kept training all of them.

If I stop the training and then resume it, it correctly loads the latest embeddings from the backup, the issue is only present if I already trained for all the epochs I initially set, then add more and resume training.

What did you expect would happen?

It should always load the embeddings from the backup, if they're present there

Relevant log output

No response

Output of pip freeze

No response