Update app.py

app.py

@@ -164,27 +164,24 @@ class ArgsNamespace:
 
 initialize_infer_state(ArgsNamespace())
 
-pipe = None
-
-def pre_load_model():
-    global pipe
-    print(f"⏳ Initializing Pipeline ({TRANSFORMER_VERSION})...")
-    try:
-        # Load to CPU explicitly
-        pipe = HunyuanVideo_1_5_Pipeline.create_pipeline(
-            pretrained_model_name_or_path=MODEL_DIR,
-            transformer_version=TRANSFORMER_VERSION,
-            enable_offloading=ENABLE_OFFLOADING,
-            enable_group_offloading=ENABLE_OFFLOADING,
-            transformer_dtype=DTYPE,
-            device=torch.device('cpu')
-        )
-        print("✅ Model loaded into CPU RAM.")
-    except Exception as e:
-        print(f"❌ Failed to load model: {e}")
-        import traceback
-        traceback.print_exc()
-        sys.exit(1)
+print(f"⏳ Initializing Pipeline ({TRANSFORMER_VERSION})...")
+try:
+    # Load to CPU explicitly
+    pipe = HunyuanVideo_1_5_Pipeline.create_pipeline(
+        pretrained_model_name_or_path=MODEL_DIR,
+        transformer_version=TRANSFORMER_VERSION,
+        enable_offloading=ENABLE_OFFLOADING,
+        enable_group_offloading=ENABLE_OFFLOADING,
+        transformer_dtype=DTYPE,
+        device=torch.device('cpu')
+    )
+    print("✅ Model loaded into CPU RAM.")
+except Exception as e:
+    print(f"❌ Failed to load model: {e}")
+    import traceback
+    traceback.print_exc()
+    sys.exit(1)
+pipe.to("cuda")
 
 def save_video_tensor(video_tensor, path, fps=24):
     if isinstance(video_tensor, list): video_tensor = video_tensor[0]
@@ -212,14 +209,8 @@ def generate(input_image, prompt, length, steps, shift, seed, guidance):
     print(f"🚀 Moving Pipeline to GPU... (Prompt: {prompt})")
     
     try:
-        # 1. Move Weights
-        pipe.to("cuda")
-        
-        # 2. FIX: Manually update internal device reference 
-        # (Hunyuan uses this attribute instead of .device in some places)
         pipe.execution_device = torch.device("cuda")
         
-        # 3. Run Inference
         output = pipe(
             prompt=prompt,
             height=480, width=854, aspect_ratio="16:9",
@@ -277,6 +268,5 @@ def create_ui():
     return demo
 
 if __name__ == "__main__":
-    pre_load_model()
     ui = create_ui()
     ui.queue().launch(server_name="0.0.0.0", share=True)