Initialization of HPSv3 seems to affect Accelerate training when used as a score function

[ODRLver]

Hi! Nice work!

I was using the HPSv3 as a score function to do some RL, the model was trained using the huggingface accelerate. However when I initializing the HPSv3 in the main process, I noticed that the Accelerator that I initialized before went crazy, for example, the accelerator.config is None after I initialized the HPSv3. After some debugging I noticed it may be caused from the HPSv3/hpsv3/utils/parser.py TrainingConfig. Since it inherited from the transformers.TrainingArguments? I don't really know why this happened.

It happeded after this line:

HPSv3/hpsv3/utils/parser.py

Line 144 in 5931ac4

return parser.parse_dict(args, allow_extra_keys=allow_extra_keys), config_path

I'm not sure if this is really the cause of my accelerator's abnormality. The version of any huggingface related stuff is pip from source.
And I found a simple solution to this, I remove the TrainingArguments in the TrainingConfig definition, and add these three lines into the TrainingConfig :

bf16 = True fp16 = False output_dir = 'output_models'

And it seems to work fine, the result for 2 example in assets folder is
10.8781 and 7.1383

[xilanhua12138]

Can you pull a request to solve it, great appreciation !

[ODRLver]

Can you pull a request to solve it, great appreciation !

New to the pull request process, made a simple one:

#11

Can close the issue anytime you want :)