fp32 for everything?

#8
by neph1 - opened

Considering the size of the safetensors file, it seems that both models are saved in fp32? I downloaded the previous version, an then only the vision model was fp32. Is this an unfortunate consequence of merging the models, or just an oversight?

I guess that an eventual quantization would solve the issue.

Owner

Oversight on my part. They were trained in 16 bit precision so no reason for the weights to be 32 bit here. Will try to fix soon.

Owner

Updated to fp16!

vikhyatk changed discussion status to closed

Sign up or log in to comment