Problems when training with fp16 mixed precision

[yhhhli]

Hi,

Thank you for sharing the code. When I am trying to use option1 to run the textural inversion in the Example notebook. There is error showing that vae's parameters are fp32. Setting the mixed_precision and revision to fp32 can run the code normally. Could you please check why fp16 is not working?

[YBouquet]

Hi,

Thanks a lot for the code and congrats for this publication !
I am also encountering issues with the revision and mixed_precision settings being set to fp16. Additionally, I'm receiving deprecation warnings from the diffusers package. Could you please provide the versions of the packages used in the project? I couldn't locate them in the repository. Thank you very much.

[kassellevi]

@YBouquet @yhhhli
i also have the problem with the precision issue.
i try to change the revision and mixed_precision to "fp32" as @yhhhli suggested and i got this error:

ValueError: Unknown mixed_precision mode: fp32. Choose between ['no', 'fp8', 'fp16', 'bf16']

what should i do to run the code normally?