Keep the model on the CPU at start-up (Scrolls-private 1a9b8870)

app.py

@@ -47,7 +47,9 @@ def log(message):
 def _load():
     global ENGINE, ENGINE_ERROR
     try:
-        ENGINE = engine.Engine(readers=(READER,))
+        # On ZeroGPU the GPU exists only inside gpu_run; a .to("cuda") in this thread is not intercepted and
+        # kills start-up, so the model stays on the CPU here and Engine.run moves it to the call's device.
+        ENGINE = engine.Engine(readers=(READER,), place_on_cuda=False)
         log(f"{READER} ready in {ENGINE.startup_s} s")
     except Exception as exc:  # noqa: BLE001
         ENGINE_ERROR = f"{type(exc).__name__}: {exc}"