How to reduce the size of finetuned model?

[cahya-wirawan]

Hi, I finetuned a medium size GPT2 model (which is around 1.35GB), but the finetuned model is around 4 times bigger (5GB), is it possible to reduce it to only the necessary parameters? Thanks

[klshuster]

you can use parlai vacuum --model-file /path/to/model to "vacuum" the model (i.e., remove the saved optimizer states)

[cahya-wirawan]

Great, thanks, I will try it