Clean up launch (remove diagnostics); document CPU-for-stub / ZeroGPU-for-real-model

server.py

@@ -24,6 +24,7 @@ from __future__ import annotations
 
 import os
 import tempfile
+import threading
 import time
 import uuid
 from collections import OrderedDict
@@ -338,36 +339,18 @@ async def index():
 
 
 # HF runs `python server.py` (README app_file) and proxies port 7860. `gradio.Server`
-# launches at MODULE TOP LEVEL; on a Space gradio binds 0.0.0.0:7860 automatically
-# and launch() blocks to keep this single long-lived server alive. This mirrors
-# gradio's reference `Server` example, which runs the same way on a ZeroGPU Space.
-# (Locally, set GRADIO_SERVER_PORT to use a different port.)
+# launches at MODULE TOP LEVEL; on a Space gradio binds 0.0.0.0:7860 automatically.
+# We launch non-blocking and hold the main thread ourselves so the (daemon) uvicorn
+# server keeps serving. (Locally, set GRADIO_SERVER_PORT to use a different port.)
 #
-# IMPORTANT: gradio is pinned to 6.10.0 on purpose. On newer gradio (6.17.x) a
-# custom gradio.Server breaks under the Space's hot-reloader — the app prints
-# "Running on 7860" but the process exits and the Space dies with RUNTIME_ERROR.
-# 6.10.0 is the version gradio's own ZeroGPU Server example ships, and it still
-# allows huggingface-hub<1.0 (needed by the real model). Keep the README
-# sdk_version and requirements*.txt gradio pins in lockstep.
-import sys as _sys
-
-
-def _log(msg):
-    print(f"[recall] {msg}", flush=True)
-    _sys.stderr.flush()
-
-
-_log("before launch")
-try:
-    _ret = app.launch(show_error=True, prevent_thread_lock=True)
-    _log(f"launch returned: {type(_ret)}")
-except Exception as _e:  # noqa: BLE001
-    import traceback
-
-    traceback.print_exc()
-    _log(f"launch raised: {type(_e).__name__}: {_e}")
-
-_log("holding main thread open")
-import threading as _threading
-
-_threading.Event().wait()
+# Two deploy gotchas, both learned the hard way:
+#   * gradio is pinned to 6.10.0. On 6.17.x a custom gradio.Server doesn't stay up
+#     under the Space runtime. 6.10.0 is the version gradio's own `Server` example
+#     ships, and it still allows huggingface-hub<1.0 (needed by the real model).
+#   * Stub mode registers no `@spaces.GPU` function, which makes a ZeroGPU Space's
+#     `spaces` reload server error out and flip the Space to RUNTIME_ERROR even
+#     though uvicorn serves fine. So the stub demo runs on CPU-basic hardware. Once
+#     the real model (with `@spaces.GPU`) is wired, switch hardware back to ZeroGPU.
+# Keep the README sdk_version and requirements*.txt gradio pins in lockstep.
+app.launch(show_error=True, prevent_thread_lock=True)
+threading.Event().wait()