sequential cpu offload

app.py

@@ -513,9 +513,8 @@ 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
     
-    # Keep everything on GPU to avoid offload hooks that break ZeroGPU pickling
-    # and avoid qfloat8 wrapping (which introduces lambda hooks) to stay pickle-safe.
-    GPU_memory_mode = "model_full_load"
+    # Use CPU offload to reduce GPU memory footprint in Space
+    GPU_memory_mode = "sequential_cpu_offload"
     compile_dit = False
     weight_dtype = torch.bfloat16
     server_name = "0.0.0.0"