Worker re-use

app.py

@@ -57,10 +57,14 @@ weights = {name: tensor.to('cuda') for name, tensor in weights.items()}
 
 del model
 
+compiled_model: AOTICompiledModel | None = None
+
 @spaces.GPU
 def run_model():
-    compiled_model: AOTICompiledModel = torch._inductor.aoti_load_package(package_path)
-    compiled_model.load_constants(weights, check_full_update=True, user_managed=True)
+    global compiled_model
+    if compiled_model is None:
+        compiled_model = torch._inductor.aoti_load_package(package_path)
+        compiled_model.load_constants(weights, check_full_update=True, user_managed=True)
     with torch.inference_mode():
         return str(compiled_model(example_inputs))