From dc8ffdffd59101f07d139bed3bcffd6435b6c004 Mon Sep 17 00:00:00 2001
From: Ben Younes <2910651+ousamabenyounes@users.noreply.github.com>
Date: Sun, 6 Sep 2026 19:11:53 +0000
Subject: [PATCH] fix(app): warm up model at startup to avoid first-request
 connection errors

---
 app.py                   |  6 +++
 tests/test_app_warmup.py | 85 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 91 insertions(+)
 create mode 100644 tests/test_app_warmup.py

diff --git a/app.py b/app.py
index 15448281..d67e5d49 100644
--- a/app.py
+++ b/app.py
@@ -563,6 +563,12 @@ def run_demo(
 ):
     demo = VoxCPMDemo(model_id=model_id, device=device)
     interface = create_demo_interface(demo)
+    # Warm up the model at startup instead of lazily on the first request.
+    # A multi-GB from_pretrained running inside the first click blocks that
+    # request for minutes and surfaces any load/download failure as an opaque
+    # per-request connection error; loading here fails fast and keeps the first
+    # generation responsive.
+    demo.get_or_load_voxcpm()
     interface.queue(max_size=10, default_concurrency_limit=1).launch(
         server_name=server_name,
         server_port=server_port,
diff --git a/tests/test_app_warmup.py b/tests/test_app_warmup.py
new file mode 100644
index 00000000..215d7227
--- /dev/null
+++ b/tests/test_app_warmup.py
@@ -0,0 +1,85 @@
+"""Regression test for issue #367: the web demo must warm up the model at
+server startup, not lazily on the first request.
+
+Loading the multi-GB model inside the first click blocks that request for
+minutes and surfaces any load/download failure as an opaque connection error.
+``run_demo`` must therefore call ``get_or_load_voxcpm()`` *before* the Gradio
+server is launched. This test records the call order and fails if the warm-up
+is removed.
+
+Heavy deps (gradio/funasr/voxcpm/torch) are not importable in the test env, so
+they are stubbed before importing ``app`` — same approach as test_cli.py.
+"""
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+from unittest import mock
+
+ROOT = Path(__file__).resolve().parents[1]
+APP_PATH = ROOT / "app.py"
+
+
+def _install_stub_deps() -> None:
+    # --- fake gradio (only the symbols app.py touches at import time) ---
+    gr = types.ModuleType("gradio")
+
+    def _passthrough(*args, **kwargs):
+        return mock.MagicMock()
+
+    gr.I18n = _passthrough
+    themes = types.SimpleNamespace(Soft=_passthrough, GoogleFont=_passthrough)
+    gr.themes = themes
+    sys.modules["gradio"] = gr
+
+    # --- fake funasr ---
+    funasr = types.ModuleType("funasr")
+    funasr.AutoModel = mock.MagicMock()
+    sys.modules["funasr"] = funasr
+
+    # --- fake voxcpm + voxcpm.model.utils ---
+    voxcpm = types.ModuleType("voxcpm")
+    voxcpm.VoxCPM = mock.MagicMock()
+    model_pkg = types.ModuleType("voxcpm.model")
+    utils = types.ModuleType("voxcpm.model.utils")
+    utils.resolve_runtime_device = lambda device, fallback: "cpu"
+    model_pkg.utils = utils
+    voxcpm.model = model_pkg
+    sys.modules["voxcpm"] = voxcpm
+    sys.modules["voxcpm.model"] = model_pkg
+    sys.modules["voxcpm.model.utils"] = utils
+
+
+def _load_app_module():
+    _install_stub_deps()
+    spec = importlib.util.spec_from_file_location("voxcpm_app_under_test", APP_PATH)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_model_is_warmed_up_before_launch():
+    app = _load_app_module()
+
+    calls: list[str] = []
+
+    class _FakeInterface:
+        def queue(self, *args, **kwargs):
+            return self
+
+        def launch(self, *args, **kwargs):
+            calls.append("launch")
+
+    with mock.patch.object(app, "create_demo_interface", return_value=_FakeInterface()), \
+         mock.patch.object(app.VoxCPMDemo, "get_or_load_voxcpm",
+                           autospec=True,
+                           side_effect=lambda self: calls.append("warmup")):
+        app.run_demo(device="cpu")
+
+    print(f"call order: {calls}")
+    assert "warmup" in calls, "model was never warmed up (loads lazily on first request)"
+    assert calls.index("warmup") < calls.index("launch"), (
+        f"model must load before launch, got order: {calls}"
+    )