Update app.py

app.py

@@ -10,18 +10,19 @@ 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
 
-# Load the Stable Diffusion pipeline and move it to the appropriate device.
+# Preload the Stable Diffusion pipeline on CPU at startup.
 pipe = DiffusionPipeline.from_pretrained(model_repo_id, torch_dtype=torch_dtype)
-pipe = pipe.to(device)
+pipe = pipe.to("cpu")
 
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 1024
 
-# Define a helper function to truncate text to a maximum of 77 tokens.
 def truncate_text(text, max_tokens=77):
+    """
+    Truncate a given text to a maximum of max_tokens using the pipeline's tokenizer.
+    """
     if text.strip() == "":
         return text
-    # Use the pipeline's tokenizer (CLIP tokenizer)
     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
@@ -40,16 +41,18 @@ def infer(
 ):
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator(device=device).manual_seed(seed)
+    generator = torch.Generator(device="cuda").manual_seed(seed)
 
-    # Truncate both prompt and negative prompt to 77 tokens.
+    # Truncate prompts to avoid CLIP token length issues.
     prompt = truncate_text(prompt, max_tokens=77)
     negative_prompt = truncate_text(negative_prompt, max_tokens=77) if negative_prompt.strip() else ""
 
-    # Move model to GPU before inference.
-    pipe.model.to("cuda")
-    
-    # Generate image using the truncated prompts.
+    # Move pipeline components to GPU
+    pipe.unet.to("cuda")
+    pipe.text_encoder.to("cuda")
+    pipe.vae.to("cuda")
+
+    # Generate the image
     image = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
@@ -60,12 +63,14 @@ def infer(
         generator=generator,
     ).images[0]
 
-    # Move model back to CPU after inference.
-    pipe.model.to("cpu")
+    # Move pipeline components back to CPU
+    pipe.unet.to("cpu")
+    pipe.text_encoder.to("cpu")
+    pipe.vae.to("cpu")
 
     return image, seed
 
-# UI layout remains unchanged.
+# Gradio UI definition
 examples = [
     "A capybara wearing a suit holding a sign that reads Hello World",
 ]
@@ -154,5 +159,4 @@ with gr.Blocks(css=css) as demo:
     )
 
 if __name__ == "__main__":
-    demo.launch(share=True)
-    
+    demo.launch()