[WIP] Don't merge / CPU offload example

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -500,6 +500,7 @@ class StableDiffusionPipeline(DiffusionPipeline):
         text_embeddings = self._encode_prompt(
             prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
         )
+        torch.cuda.empty_cache()
 
         # 4. Prepare timesteps
         self.scheduler.set_timesteps(num_inference_steps, device=device)