Textual inv save log memory (#2184)

* Quality check and adding tokenizer

* Adapted stable diffusion to mixed precision+finished up style fixes

* Fixed based on patrick's review

* Fixed oom from number of validation images

* Removed unnecessary np.array conversion

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

examples/textual_inversion/textual_inversion.py

@@ -781,7 +781,10 @@ def main():
                 args.pretrained_model_name_or_path,
                 text_encoder=accelerator.unwrap_model(text_encoder),
                 tokenizer=tokenizer,
+                unet=unet,
+                vae=vae,
                 revision=args.revision,
+                torch_dtype=weight_dtype,
             )
             pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)
             pipeline = pipeline.to(accelerator.device)
@@ -791,8 +794,11 @@ def main():
             generator = (
                 None if args.seed is None else torch.Generator(device=accelerator.device).manual_seed(args.seed)
             )
-            prompt = args.num_validation_images * [args.validation_prompt]
-            images = pipeline(prompt, num_inference_steps=25, generator=generator).images
+            images = []
+            for _ in range(args.num_validation_images):
+                with torch.autocast("cuda"):
+                    image = pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
+                images.append(image)
 
             for tracker in accelerator.trackers:
                 if tracker.name == "tensorboard":