[UNet2DConditionModel, UNet2DModel] pass norm_num_groups to all the blocks #442
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
|
Thank you everyone! |
patrickvonplaten
left a comment
patrickvonplaten
left a comment
There was a problem hiding this comment.
Very cool @patil-suraj !
|
|
||
| self.assertTrue(torch.allclose(output_slice, expected_output_slice, rtol=1e-2)) | ||
|
|
||
| def test_forward_with_norm_groups(self): |
There was a problem hiding this comment.
It seems to adjust norm groups inside the blocks cf https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_blocks.py#L780
So not sure if it can work with any value because of the division with 4
patrickvonplaten
left a comment
patrickvonplaten
left a comment
There was a problem hiding this comment.
Sorry @patil-suraj - let's really try to not change existing test (except for they are wrong). It would be really nice to just add a new test here
Agree, I have already added a new common test |
…locks (huggingface#442) * pass norm_num_groups to unet blocs and attention * fix UNet2DConditionModel * add norm_num_groups arg in vae * add tests * remove comment * Apply suggestions from code review
This PR removed the hardcoded value of 32 for
norm_num_groupsand makes sure that the arg is passed to all up, down, res, and attention blocks.Fixes #410