Why are the V2 weights twice the size as V3?

#102
by stri8ted - opened

In v2, model.safetensors is 6.17 GB. In v3, the size is 3.09 GB.

is it because v3 weights are in float32, and b4 is float16?

If yes, is the float16 version published anywhere? Or, does the transformer libs provide a way of converting it to float16?

Thanks.

Sign up or log in to comment