[SVD] support generators that are created on GPU (huggingface#6484)

* debug generator

* fix?

* fix?

* fix

* remove print.

* revert none check

Author: sayakpaul

src/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py

@@ -429,15 +429,20 @@ def __call__(
         fps = fps - 1
 
         # 4. Encode input image using VAE
-        image = self.image_processor.preprocess(image, height=height, width=width)
-        noise = randn_tensor(image.shape, generator=generator, device=image.device, dtype=image.dtype)
+        image = self.image_processor.preprocess(image, height=height, width=width).to(device)
+        noise = randn_tensor(image.shape, generator=generator, device=device, dtype=image.dtype)
         image = image + noise_aug_strength * noise
 
         needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
         if needs_upcasting:
             self.vae.to(dtype=torch.float32)
 
-        image_latents = self._encode_vae_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance)
+        image_latents = self._encode_vae_image(
+            image,
+            device=device,
+            num_videos_per_prompt=num_videos_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+        )
         image_latents = image_latents.to(image_embeddings.dtype)
 
         # cast back to fp16 if needed