`AsymmetricAutoencoderKL`: missing `generator` argument in `decode()` called from `forward()`

[townwish4git]

Describe the bug

AsymmetricAutoencoderKL.decode() has parameter generator,

def decode(
        self,
        z: torch.FloatTensor,
        generator: Optional[torch.Generator] = None,
        image: Optional[torch.FloatTensor] = None,
        mask: Optional[torch.FloatTensor] = None,
        return_dict: bool = True,
    ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
    ...

But when forward() calls self.decode(), only three positional arguments are passed in,

...
dec = self.decode(z, sample, mask).sample
...

which means:

1. argument z => parameter z;
2. argument sample => parameter generator;
3. argument mask => parameter image;
4. default None => parameter mask

it appears to be a bug and should be corrected to

- dec = self.decode(z, sample, mask).sample
+ dec = self.decode(z, generator, sample, mask).sample

Reproduction

from diffusers import AsymmetricAutoencoderKL

...

Logs

No response

System Info

Environment-independent

Who can help?

@sayakpaul @DN6 @yiyixuxu @cross-attention

[sayakpaul]

Thanks for opening this issue. Would you like to take a stab at opening a PR to fix this?

[townwish4git]

Thanks for opening this issue. Would you like to take a stab at opening a PR to fix this?

Yes of course. Here is #8378