Merge pull request #425 from pumpkinperson996/agent/fix-whisper-model-selection

Fix Whisper model selection caching

backend/app/services/note.py

@@ -262,7 +262,10 @@ class NoteGenerator:
             raise Exception(f"不支持的转写器：{self.transcriber_type}")
 
         logger.info(f"使用转写器：{self.transcriber_type}")
-        return get_transcriber(transcriber_type=self.transcriber_type)
+        return get_transcriber(
+            transcriber_type=self.transcriber_type,
+            model_size=self.model_size,
+        )
 
     def _get_gpt(self, model_name: Optional[str], provider_id: Optional[str]) -> GPT:
         """

backend/app/transcriber/transcriber_provider.py

@@ -1,5 +1,6 @@
 import os
 import platform
+import threading
 from enum import Enum
 
 from app.transcriber.groq import GroqTranscriber
@@ -38,17 +39,29 @@ _transcribers = {
     TranscriberType.GROQ: None,
 }
 
+# Cache instances together with their constructor configuration. The
+# transcriber choice and Whisper model size can be changed from the frontend,
+# so caching by transcriber type alone would keep using the first loaded model.
+_transcriber_configs = {key: None for key in _transcribers}
+_transcriber_init_lock = threading.Lock()
+
 # 公共实例初始化函数
 def _init_transcriber(key: TranscriberType, cls, *args, **kwargs):
-    if _transcribers[key] is None:
+    init_config = (args, tuple(sorted(kwargs.items())))
-        logger.info(f'创建 {cls.__name__} 实例: {key}')
+    with _transcriber_init_lock:
-        try:
+        instance = _transcribers[key]
-            _transcribers[key] = cls(*args, **kwargs)
+        if instance is None or _transcriber_configs[key] != init_config:
+            action = "创建" if instance is None else "按新配置重新创建"
+            logger.info(f'{action} {cls.__name__} 实例: {key}')
+            try:
+                new_instance = cls(*args, **kwargs)
+            except Exception as e:
+                logger.error(f"{cls.__name__} 创建失败: {e}")
+                raise
+            _transcribers[key] = new_instance
+            _transcriber_configs[key] = init_config
             logger.info(f'{cls.__name__} 创建成功')
-        except Exception as e:
+        return _transcribers[key]
-            logger.error(f"{cls.__name__} 创建失败: {e}")
-            raise
-    return _transcribers[key]
 
 # 各类型获取方法
 def get_groq_transcriber():
@@ -70,13 +83,13 @@ def get_mlx_whisper_transcriber(model_size="base"):
     return _init_transcriber(TranscriberType.MLX_WHISPER, MLXWhisperTranscriber, model_size=model_size)
 
 # 通用入口
-def get_transcriber(transcriber_type="fast-whisper", model_size="base", device="cuda"):
+def get_transcriber(transcriber_type="fast-whisper", model_size=None, device="cuda"):
     """
     获取指定类型的转录器实例
 
     参数:
         transcriber_type: 支持 "fast-whisper", "mlx-whisper", "bcut", "kuaishou", "groq"
-        model_size: 模型大小，适用于 whisper 类
+        model_size: 模型大小，适用于 whisper 类；未提供时才读取环境变量默认值
         device: 设备类型（如 cuda / cpu），仅 whisper 使用
 
     返回:
@@ -90,7 +103,9 @@ def get_transcriber(transcriber_type="fast-whisper", model_size="base", device="
         logger.warning(f'未知转录器类型 "{transcriber_type}"，默认使用 fast-whisper')
         transcriber_enum = TranscriberType.FAST_WHISPER
 
-    whisper_model_size = os.environ.get("WHISPER_MODEL_SIZE", model_size)
+    # The explicit value normally comes from the persisted frontend setting and
+    # must take precedence over Docker's startup default.
+    whisper_model_size = model_size or os.environ.get("WHISPER_MODEL_SIZE", "base")
 
     if transcriber_enum == TranscriberType.FAST_WHISPER:
         return get_whisper_transcriber(whisper_model_size, device=device)

backend/tests/test_whisper_config_forwarding.py

@@ -0,0 +1,120 @@
+import os
+import pathlib
+import subprocess
+import sys
+import textwrap
+
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+
+
+def _run_isolated(script: str) -> None:
+    env = os.environ.copy()
+    env["PYTHONPATH"] = os.pathsep.join(
+        filter(None, [str(ROOT), env.get("PYTHONPATH")])
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", textwrap.dedent(script)],
+        cwd=ROOT,
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+    assert result.returncode == 0, result.stdout + result.stderr
+
+
+def test_note_generator_forwards_configured_whisper_model_size():
+    _run_isolated(
+        """
+        from app.services import note as note_service
+
+        calls = []
+
+        def fake_get_transcriber(**kwargs):
+            calls.append(kwargs)
+            return object()
+
+        note_service.get_transcriber = fake_get_transcriber
+
+        generator = note_service.NoteGenerator.__new__(note_service.NoteGenerator)
+        generator.transcriber_type = "fast-whisper"
+        generator.model_size = "large-v3-turbo"
+
+        generator._init_transcriber()
+
+        assert calls == [
+            {
+                "transcriber_type": "fast-whisper",
+                "model_size": "large-v3-turbo",
+            }
+        ]
+        """
+    )
+
+
+def test_whisper_cache_is_rebuilt_when_model_size_changes():
+    _run_isolated(
+        """
+        from app.transcriber import transcriber_provider as provider
+
+        class FakeWhisperTranscriber:
+            def __init__(self, model_size, device):
+                self.model_size = model_size
+                self.device = device
+
+        provider._transcribers = {key: None for key in provider._transcribers}
+        provider._transcriber_configs = {
+            key: None for key in provider._transcribers
+        }
+        provider.WhisperTranscriber = FakeWhisperTranscriber
+
+        base = provider.get_whisper_transcriber("base", device="cpu")
+        turbo = provider.get_whisper_transcriber("large-v3-turbo", device="cpu")
+        turbo_again = provider.get_whisper_transcriber(
+            "large-v3-turbo", device="cpu"
+        )
+
+        assert base.model_size == "base"
+        assert turbo.model_size == "large-v3-turbo"
+        assert turbo is not base
+        assert turbo_again is turbo
+        """
+    )
+
+
+def test_explicit_model_size_wins_over_environment_default():
+    _run_isolated(
+        """
+        import os
+
+        from app.transcriber import transcriber_provider as provider
+
+        class FakeWhisperTranscriber:
+            def __init__(self, model_size, device):
+                self.model_size = model_size
+                self.device = device
+
+        os.environ["WHISPER_MODEL_SIZE"] = "tiny"
+        provider._transcribers = {key: None for key in provider._transcribers}
+        provider._transcriber_configs = {
+            key: None for key in provider._transcribers
+        }
+        provider.WhisperTranscriber = FakeWhisperTranscriber
+
+        transcriber = provider.get_transcriber(
+            transcriber_type="fast-whisper",
+            model_size="large-v3-turbo",
+            device="cpu",
+        )
+
+        assert transcriber.model_size == "large-v3-turbo"
+
+        fallback = provider.get_transcriber(
+            transcriber_type="fast-whisper",
+            model_size=None,
+            device="cpu",
+        )
+
+        assert fallback.model_size == "tiny"
+        """
+    )