Correction for non-integral image resolutions with quantizations other than float32 (#7356)

* Correction for non-integral image resolutions with quantizations other than float32.

* Support for training, and use of diffusers-style casting.

src/diffusers/models/unets/unet_stable_cascade.py

@@ -521,9 +521,11 @@ class StableCascadeUNet(ModelMixin, ConfigMixin, FromOriginalUNetMixin):
                         if isinstance(block, SDCascadeResBlock):
                             skip = level_outputs[i] if k == 0 and i > 0 else None
                             if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+                                orig_type = x.dtype
                                 x = torch.nn.functional.interpolate(
                                     x.float(), skip.shape[-2:], mode="bilinear", align_corners=True
                                 )
+                                x = x.to(orig_type)
                             x = torch.utils.checkpoint.checkpoint(
                                 create_custom_forward(block), x, skip, use_reentrant=False
                             )
@@ -547,9 +549,11 @@ class StableCascadeUNet(ModelMixin, ConfigMixin, FromOriginalUNetMixin):
                         if isinstance(block, SDCascadeResBlock):
                             skip = level_outputs[i] if k == 0 and i > 0 else None
                             if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+                                orig_type = x.dtype
                                 x = torch.nn.functional.interpolate(
                                     x.float(), skip.shape[-2:], mode="bilinear", align_corners=True
                                 )
+                                x = x.to(orig_type)
                             x = block(x, skip)
                         elif isinstance(block, SDCascadeAttnBlock):
                             x = block(x, clip)