Avoiding graph break by changing the way we infer dtype in vae.decoder (#12512)

* Changing the way we infer dtype to avoid force evaluation of lazy tensors

* changing way to infer dtype to ensure type consistency

* more robust infering of dtype

* removing the upscale dtype entirely

src/diffusers/models/autoencoders/vae.py

@@ -286,11 +286,9 @@ class Decoder(nn.Module):
 
         sample = self.conv_in(sample)
 
-        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
         if torch.is_grad_enabled() and self.gradient_checkpointing:
             # middle
             sample = self._gradient_checkpointing_func(self.mid_block, sample, latent_embeds)
-            sample = sample.to(upscale_dtype)
 
             # up
             for up_block in self.up_blocks:
@@ -298,7 +296,6 @@ class Decoder(nn.Module):
         else:
             # middle
             sample = self.mid_block(sample, latent_embeds)
-            sample = sample.to(upscale_dtype)
 
             # up
             for up_block in self.up_blocks: