The uploaded vae safetensor model in fp16

by qiwang - opened

some parameters of vae in fp16 version are stored in float32 mode so that i got an error while runing official code.
unmatching information: Input type (torch.cuda.HalfTensor) and weight type (torch.cuda.FloatTensor)

Solved: the cause of this bug is that i use a conda envrionment with python3.7 and torch1.13, which is incompatible with sdxl.

Sign up or log in to comment