reload model

app.py

@@ -354,14 +354,15 @@ def ui(GPU_memory_mode, scheduler_dict, config_path, compile_dit, weight_dtype):
         
         with gr.Column(variant="panel"):
             # Hide model selection
-            diffusion_transformer_dropdown, _ = create_model_checkpoints(controller, visible=False, default_model="Wan-AI/Wan2.1-T2V-14B")
+            local_model_dir = os.path.join("models", "Wan2.1-T2V-14B")
+            diffusion_transformer_dropdown, _ = create_model_checkpoints(controller, visible=False, default_model=local_model_dir)
             
             # Use snapshot download for the VideoCoF repo to get all weights (including safetensors)
             try:
                 from huggingface_hub import snapshot_download, hf_hub_download
                 print("Downloading Wan2.1-T2V-14B weights...")
-                default_model_path = "Wan-AI/Wan2.1-T2V-14B"
-                snapshot_download(repo_id=default_model_path, local_dir=default_model_path)
+                hf_model_id = "Wan-AI/Wan2.1-T2V-14B"
+                snapshot_download(repo_id=hf_model_id, local_dir=local_model_dir, local_dir_use_symlinks=False)
                 
                 os.makedirs("models/Personalized_Model", exist_ok=True)
                 
@@ -385,7 +386,7 @@ def ui(GPU_memory_mode, scheduler_dict, config_path, compile_dit, weight_dtype):
 
         # Preload heavy weights and LoRAs before launching the UI to avoid first-run latency.
         acc_lora_path = os.path.join("models", "Personalized_Model", "Wan2.1_Text_to_Video_14B_FusionX_LoRA.safetensors")
-        preload_models(controller, default_model_path, "videocof.safetensors", acc_lora_path)
+        preload_models(controller, local_model_dir, "videocof.safetensors", acc_lora_path)
 
         with gr.Column(variant="panel"):
             prompt_textbox, negative_prompt_textbox = create_prompts(prompt="Remove the young man with short black hair wearing black shirt on the left.")
@@ -498,7 +499,8 @@ def ui(GPU_memory_mode, scheduler_dict, config_path, compile_dit, weight_dtype):
 if __name__ == "__main__":
     from videox_fun.ui.controller import flow_scheduler_dict
     
-    GPU_memory_mode = "sequential_cpu_offload"
+    # Keep everything on GPU to avoid offload hooks that break ZeroGPU pickling
+    GPU_memory_mode = "model_full_load_and_qfloat8"
     compile_dit = False
     weight_dtype = torch.bfloat16
     server_name = "0.0.0.0"