This compiles the repeated transformer blocks for ~2x faster initialization.

app.py

@@ -5,14 +5,13 @@ 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):
@@ -28,12 +27,16 @@ 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
+# Load Qwen/Qwen-Image pipeline with regional compilation
 dtype = torch.bfloat16
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-# --- Model Loading ---
-pipe_qwen = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=dtype).to(device)
+# --- Model Loading with Regional Compilation ---
+ckpt_id = "Qwen/Qwen-Image"
+pipe_qwen = DiffusionPipeline.from_pretrained(
+    ckpt_id, torch_dtype=dtype
+).to(device)
+pipe_qwen.transformer.compile_repeated_blocks(fullgraph=True)
 
 # Aspect ratios
 aspect_ratios = {