Closed Thytu closed 5 months ago
It appears that Trainer
currently does not provide a way to add metadata to the saved model.safetensors
file.
A work-around would be to load the saved model.safetensors
file, add the metadata, and re-save it again. It will add some processing time at the end of each save_steps
but it will result in a more friendly way to load a train model.
A user will be able to simply provide a tensor tensor and SMIT
could automatically adapt the model (LoRA / Quant) depending the config file provided.
As a fallback in case the safetensor
file does not contain the config file in its metadata, SMIT should also provide a method that will takes as input both the safetensor
and a config as dict/file.
The current process of loading a trained model can be quite challenging for newcomers (cf), this can be both frustrating and discouraging for them.
SMIT should provide a
classmethod
that would automatically load a trained model given a path to asafe_tensor
and the config file used to generate the model.Ultimately it would great to not have to provide the config file by maybe storing everything into the
safe_tensor
file.