修复直播测试与 TTS 预热

tests/test_faster_qwen_worker.py

@@ -9,6 +9,7 @@ from app.faster_qwen_worker import (
     MAX_NEW_TOKENS,
     FasterQwenWorkerClient,
     FasterQwenWorkerTimeout,
+    faster_qwen_worker_main,
     _generation_kwargs,
     _configure_worker_environment,
 )
@@ -91,6 +92,14 @@ def _ready(pid):
     }
 
 
+def _progress(stage):
+    return {
+        "type": "progress",
+        "stage": stage,
+        "pid": 4000,
+    }
+
+
 def _probe_worker_main(connection, _settings):
     connection.send({
         "type": "ready",
@@ -153,6 +162,8 @@ class FasterQwenWorkerTests(unittest.TestCase):
 
     def test_client_returns_worker_audio(self):
         connection = _FakeConnection([
+            _progress("model_loading"),
+            _progress("warmup_started"),
             _ready(4100),
             {
                 "type": "result",
@@ -181,6 +192,81 @@ class FasterQwenWorkerTests(unittest.TestCase):
         self.assertEqual(len(context.processes), 1)
         client.close()
 
+    def test_client_default_startup_timeout_allows_slow_model_load(self):
+        client = FasterQwenWorkerClient({})
+
+        self.assertEqual(client.startup_timeout_seconds, 900.0)
+        client.close()
+
+    def test_worker_applies_process_limits_after_model_warmup(self):
+        order = []
+
+        class FakeModel:
+            def warmup(self):
+                order.append("warmup")
+
+        class WorkerConnection:
+            def __init__(self):
+                self.sent = []
+                self.closed = False
+
+            def send(self, message):
+                self.sent.append(message)
+
+            def recv(self):
+                return {"command": "stop"}
+
+            def close(self):
+                self.closed = True
+
+        connection = WorkerConnection()
+        with (
+            patch("app.faster_qwen_worker._configure_worker_environment", return_value=3),
+            patch(
+                "app.faster_qwen_worker._load_runtime",
+                side_effect=lambda _settings, progress=None: (
+                    progress("from_pretrained_started") if progress else None,
+                    order.append("load"),
+                    {"model": FakeModel()},
+                )[-1],
+            ),
+            patch(
+                "app.faster_qwen_worker._apply_worker_process_limits",
+                side_effect=lambda _settings: order.append("limits") or {
+                    "cpu_affinity_count": 6,
+                    "process_priority": "below_normal",
+                },
+            ),
+            patch(
+                "app.faster_qwen_worker._synthesize_wav",
+                side_effect=lambda _runtime, _text: order.append("voice_clone") or b"wav",
+            ),
+        ):
+            faster_qwen_worker_main(connection, {})
+
+        self.assertEqual(order, ["load", "warmup", "voice_clone", "limits"])
+        progress_stages = [
+            item.get("stage")
+            for item in connection.sent
+            if item.get("type") == "progress"
+        ]
+        self.assertEqual(
+            progress_stages,
+            [
+                "process_started",
+                "environment_ready",
+                "model_loading",
+                "from_pretrained_started",
+                "model_loaded",
+                "warmup_started",
+                "graph_warmup_complete",
+                "voice_clone_warmup_started",
+                "voice_clone_warmup_complete",
+                "warmup_complete",
+            ],
+        )
+        self.assertEqual(connection.sent[-1]["type"], "ready")
+
     def test_timeout_terminates_worker_and_starts_a_prewarmed_replacement(self):
         first_connection = _FakeConnection([_ready(4200)])
         replacement_connection = _FakeConnection([_ready(4300)])