ZeroGPU: load student on CPU at startup with cuda.is_available guarded False

server_app.py

@@ -36,11 +36,13 @@ def viz() -> dict:
 @app.api(name="probe", stream_every=0.1)
 @spaces.GPU
 def probe(text: str) -> dict:  # annotation = type of each streamed chunk
-    if not text.strip() or not RT.model_ready:
-        yield {"error": "model not ready" if text.strip() else "empty prompt",
-               "done": True}
+    if not text.strip():
+        yield {"error": "empty prompt", "done": True}
+        return
+    _boot.to_device(RT)   # lazy-loads the student here (real CUDA inside @spaces.GPU)
+    if not RT.model_ready:
+        yield {"error": RT.load_error or "model not ready", "done": True}
         return
-    _boot.to_device(RT)
     partial, gates = "", []
     for partial, gates in _boot.stream(RT, text):
         yield {"text": partial, "gates": gates, "done": False}
@@ -54,11 +56,13 @@ def probe(text: str) -> dict:  # annotation = type of each streamed chunk
 @app.api(name="arena", stream_every=0.1)
 @spaces.GPU
 def arena(text: str) -> dict:  # annotation = type of each streamed chunk
-    if not text.strip() or not RT.model_ready:
-        yield {"error": "model not ready" if text.strip() else "empty prompt",
-               "done": True}
+    if not text.strip():
+        yield {"error": "empty prompt", "done": True}
+        return
+    _boot.to_device(RT)   # lazy-loads the student here (real CUDA inside @spaces.GPU)
+    if not RT.model_ready:
+        yield {"error": RT.load_error or "model not ready", "done": True}
         return
-    _boot.to_device(RT)
     base = deku = ""
     gates: list[float] = []
     for base, deku, gates in _boot.stream_pair(RT, text):