Update app.py

app.py

@@ -5,21 +5,24 @@ from PIL import Image
 from diffusers import DiffusionPipeline
 import random
 import uuid
+from typing import Union, List, Optional
 import numpy as np
 import time
 import zipfile
 import os
 
 # Description for the app
-DESCRIPTION = """## Qwen Image HPC/."""
+DESCRIPTION = """## Qwen Image Hpc/."""
 
 # Helper functions
 def save_image(img):
+    """Saves a PIL image to a file with a unique name."""
     unique_name = str(uuid.uuid4()) + ".png"
     img.save(unique_name)
     return unique_name
 
 def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    """Generates a random seed if the randomize option is enabled."""
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
     return seed
@@ -27,24 +30,32 @@ def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
 MAX_SEED = np.iinfo(np.int32).max
 MAX_IMAGE_SIZE = 2048
 
-# Load Qwen/Qwen-Image pipeline with regional compilation
+# Load Qwen/Qwen-Image pipeline
 dtype = torch.bfloat16
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-# --- Model Loading with Regional Compilation ---
-ckpt_id = "Qwen/Qwen-Image"
+# --- Model Loading ---
+# Load the pipeline from pretrained weights
 pipe_qwen = DiffusionPipeline.from_pretrained(
-    ckpt_id, torch_dtype=dtype
+    "Qwen/Qwen-Image",
+    torch_dtype=dtype
 ).to(device)
-pipe_qwen.transformer.compile_repeated_blocks(fullgraph=True)
+
+# --- Regional Compilation ---
+# Apply regional compilation to speed up cold-starts while retaining benefits.
+# This compiles the repeated transformer blocks for ~2x faster initialization.
+print("Applying regional compilation...")
+pipe_qwen.transformer.compile(fullgraph=True, mode="reduce-overhead")
+print("Compilation complete.")
+
 
 # Aspect ratios
 aspect_ratios = {
-    "1:1": (1328, 1328),
-    "16:9": (1664, 928),
-    "9:16": (928, 1664),
-    "4:3": (1472, 1140),
-    "3:4": (1140, 1472)
+    "1:1": (1024, 1024),
+    "16:9": (1344, 768),
+    "9:16": (768, 1344),
+    "4:3": (1152, 896),
+    "3:4": (896, 1152)
 }
 
 # Generation function for Qwen/Qwen-Image
@@ -281,4 +292,4 @@ with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
     )
 
 if __name__ == "__main__":
-    demo.queue(max_size=50).launch(share=False, mcp_server=True, ssr_mode=False, show_error=True)
+    demo.queue(max_size=50).launch(share=False)