Make UNet2DConditionOutput pickle-able

[prathikr]

This PR addresses previous concerns that the output of the UNet's forward pass is not copy-able. The root cause appears to be because copy fails on collections.OrderedDict dataclass with required args. The solution presented sets a default value for sample such that is it no longer a required parameter of the output class while still erroring when missing since the default setting is None (link to similar solution for different model).

Reproduction Instructions:

from diffusers.utils import BaseOutput
from dataclasses import dataclass
import copy

@dataclass
class NetParams(BaseOutput):
    sample: torch.FloatTensor

m = NetParams(sample=torch.randn(1, 10))
n = copy.copy(m)

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[prathikr]

@patrickvonplaten @anton-l can I please get a review on this?

[prathikr]

@patrickvonplaten @anton-l any updates?

[patrickvonplaten]

This change is ok for me! Could we add a test here that shows how we can now pickle the output?

[prathikr]

@patrickvonplaten I gave adding a unit test a try. Let me know if I should change it or put it somewhere else.

[prathikr]

@patrickvonplaten any updates?

[prathikr]

@patrickvonplaten this is currently blocking ONNX Runtime integration with Diffusers. Can you please provide an update? Thank you.

[patrickvonplaten]

Ok for me!

[patrickvonplaten]

@sayakpaul @pcuenca could you maybe also quickly check?

[sayakpaul]

Thanks for the change! PR looks great to me except for https://github.com/huggingface/diffusers/pull/3857/files#r1252260499.

[prathikr]

@sayakpaul can you please review/merge? Thank you.

[sayakpaul]

Thanks for iterating!