Avoid negative strides for tensors (#717)

* Avoid negative strides for tensors

* Changed not to make torch.tensor

* Removed a needless copy

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -304,12 +304,9 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
         latents = init_latents
 
         t_start = max(num_inference_steps - init_timestep + offset, 0)
+        timesteps = self.scheduler.timesteps[t_start:]
 
-        # Some schedulers like PNDM have timesteps as arrays
-        # It's more optimzed to move all timesteps to correct device beforehand
-        timesteps_tensor = torch.tensor(self.scheduler.timesteps[t_start:], device=self.device)
-
-        for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+        for i, t in enumerate(self.progress_bar(timesteps)):
             t_index = t_start + i
 
             # expand the latents if we are doing classifier free guidance

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -342,12 +342,9 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
         latents = init_latents
 
         t_start = max(num_inference_steps - init_timestep + offset, 0)
+        timesteps = self.scheduler.timesteps[t_start:]
 
-        # Some schedulers like PNDM have timesteps as arrays
-        # It's more optimzed to move all timesteps to correct device beforehand
-        timesteps_tensor = torch.tensor(self.scheduler.timesteps[t_start:], device=self.device)
-
-        for i, t in tqdm(enumerate(timesteps_tensor)):
+        for i, t in tqdm(enumerate(timesteps)):
             t_index = t_start + i
             # expand the latents if we are doing classifier free guidance
             latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents