Upload 127 files

app.py

@@ -24,6 +24,10 @@ import uuid
 # setdefault) because Spaces exports GRADIO_SSR_MODE=true.
 os.environ["GRADIO_SSR_MODE"] = "false"
 
+# On ZeroGPU, `spaces` must be imported before any CUDA-touching library.
+with contextlib.suppress(Exception):  # not installed in local dev
+    import spaces  # noqa: F401
+
 import gradio as gr
 from fastapi import FastAPI, WebSocket, WebSocketDisconnect
 from fastapi.responses import FileResponse
@@ -356,6 +360,21 @@ with gr.Blocks(title="HuggingWizards — Game Master Dashboard") as demo:
 
 app = gr.mount_gradio_app(app, demo, path="/dashboard", ssr_mode=False)
 
+# ---- ZeroGPU startup report ------------------------------------------------
+# The `spaces` package sends its "this app has @spaces.GPU functions" report
+# to the ZeroGPU controller via a hook it installs on gr.Blocks.launch(). We
+# serve everything through uvicorn instead of demo.launch(), so that hook
+# never fires and ZeroGPU kills the Space with "No @spaces.GPU function
+# detected during startup". Invoke the same startup task manually. It packs
+# the (already loaded) model for fast GPU transfer and phones home; off
+# Spaces, `spaces.zero` doesn't define `startup` and this is a no-op.
+with contextlib.suppress(Exception):
+    from spaces import zero as _spaces_zero  # type: ignore
+
+    if hasattr(_spaces_zero, "startup"):
+        _spaces_zero.startup()
+        print("[app] ZeroGPU startup report sent")
+
 
 if __name__ == "__main__":
     import uvicorn

game/gamemaster.py

@@ -161,14 +161,6 @@ def _ensure_model():
         return False
 
 
-# On Spaces, load the model at startup (ZeroGPU replays the .to("cuda") once a
-# real GPU is attached) so the @spaces.GPU window is spent on inference only —
-# lazy-loading inside the first GPU call would blow the 60 s duration on the
-# weight download. Locally we stay lazy so dev/CI never downloads weights.
-if _ON_SPACES and not os.environ.get("HUGGINGWIZARDS_NO_LLM"):
-    _ensure_model()
-
-
 # ---- mercy guardrail -------------------------------------------------------
 # The GM may slow the boss's attack speed to help wounded parties, but its
 # willingness to help decays as waves progress: the allowed floor rises from
@@ -209,6 +201,18 @@ def _run_model(system: str, user: str) -> str:
     return text
 
 
+# On Spaces, load the model at startup (ZeroGPU replays the .to("cuda") once a
+# real GPU is attached) so the @spaces.GPU window is spent on inference only —
+# lazy-loading inside the first GPU call would blow the 60 s duration on the
+# weight download. This MUST come after the @spaces.GPU function above is
+# defined: ZeroGPU's startup probe needs to see a registered GPU function, and
+# the multi-minute weight download would otherwise delay that registration
+# ("No @spaces.GPU function detected during startup"). Locally we stay lazy so
+# dev/CI never downloads weights.
+if _ON_SPACES and not os.environ.get("HUGGINGWIZARDS_NO_LLM"):
+    _ensure_model()
+
+
 def _extract_json(text: str) -> dict | None:
     m = re.search(r"\{.*\}", text, re.DOTALL)
     if not m: