Update app.py

app.py

@@ -10,19 +10,20 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 model_repo_id = "stabilityai/stable-diffusion-3.5-large"
 torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
 
-# Preload the Stable Diffusion pipeline on CPU at startup.
+# Preload the Stable Diffusion pipeline on GPU (if available)
 pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to("cpu")
+pipe = pipe.to(device)
 
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 1024
 
 def truncate_text(text, max_tokens=77):
     """
-    Truncate a given text to a maximum of max_tokens using the pipeline's tokenizer.
+    Explicitly truncate a given text to a maximum of `max_tokens` using the pipeline's tokenizer.
     """
     if text.strip() == "":
         return text
+    # Tokenize with truncation enabled and a maximum length
     tokens = pipe.tokenizer(text, truncation=True, max_length=max_tokens, add_special_tokens=True)
     truncated_text = pipe.tokenizer.decode(tokens["input_ids"], skip_special_tokens=True)
     return truncated_text
@@ -39,20 +40,16 @@ def infer(
     num_inference_steps=40,
     progress=gr.Progress(track_tqdm=True),
 ):
+    # Randomize seed if requested
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator(device="cuda").manual_seed(seed)
-
-    # Truncate prompts to avoid CLIP token length issues.
+    generator = torch.Generator(device=device).manual_seed(seed)
+    
+    # Explicitly truncate both prompt and negative prompt to avoid CLIP token warnings.
     prompt = truncate_text(prompt, max_tokens=77)
     negative_prompt = truncate_text(negative_prompt, max_tokens=77) if negative_prompt.strip() else ""
-
-    # Move pipeline components to GPU
-    pipe.unet.to("cuda")
-    pipe.text_encoder.to("cuda")
-    pipe.vae.to("cuda")
-
-    # Generate the image
+    
+    # Generate the image (the pipeline is already on GPU)
     image = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
@@ -62,15 +59,10 @@ def infer(
         height=height,
         generator=generator,
     ).images[0]
-
-    # Move pipeline components back to CPU
-    pipe.unet.to("cpu")
-    pipe.text_encoder.to("cpu")
-    pipe.vae.to("cpu")
-
+    
     return image, seed
 
-# Gradio UI definition
+# Gradio UI layout remains as before.
 examples = [
     "A capybara wearing a suit holding a sign that reads Hello World",
 ]