The open-source AI voice studio. Clone, dictate, create.
Extract OpenAI-compat cache key + fix misleading lock comment
Moroya Sakamotoea2cf01d5ec0
diff --git a/backend/backends/__init__.py b/backend/backends/__init__.pyindex 2f7d3a71..833bdc23 100644--- a/backend/backends/__init__.py+++ b/backend/backends/__init__.py@@ -222,6 +222,11 @@ _custom_llm_endpoint: Optional[str] = None_custom_llm_model: Optional[str] = None_custom_llm_api_key: Optional[str] = None+# ``_llm_backends`` key for the singleton OpenAI-compat backend. Extracted+# so ``set_llm_config`` (cache invalidation) and ``get_llm_backend`` (cache+# lookup + on-miss install) can't drift on the string literal.+_OPENAI_COMPAT_CACHE_KEY = "openai_compat"+# Supported TTS engines — keyed by engine name, value is the backend class import path.# The factory function uses this for the if/elif chain; the model configs live on the backend classes.TTS_ENGINES = {@@ -829,7 +834,7 @@ def set_llm_config(_custom_llm_model = model_custom_llm_api_key = api_keyif changed:- _llm_backends.pop("openai_compat", None)+ _llm_backends.pop(_OPENAI_COMPAT_CACHE_KEY, None)def get_llm_backend() -> LLMBackend:@@ -852,7 +857,7 @@ def get_llm_backend() -> LLMBackend:model = _custom_llm_modelapi_key = _custom_llm_api_keyif endpoint and model:- cached = _llm_backends.get("openai_compat")+ cached = _llm_backends.get(_OPENAI_COMPAT_CACHE_KEY)if cached is not None:return cachedbackend = OpenAICompatLLMBackend(@@ -860,13 +865,15 @@ def get_llm_backend() -> LLMBackend:model=model,api_key=api_key,)- _llm_backends["openai_compat"] = backend+ _llm_backends[_OPENAI_COMPAT_CACHE_KEY] = backendreturn backend- # No custom config — hand off to the Qwen dispatch, which has its own- # lock and can take multiple seconds to load a model. Do that outside- # ``_llm_backends_lock`` so a slow model load doesn't block concurrent- # ``set_llm_config()`` writes.+ # No custom config — hand off to the Qwen dispatch. ``get_llm_backend_for_engine``+ # reuses this exact ``_llm_backends_lock`` (see its body below), and+ # ``threading.Lock`` is non-reentrant, so this call MUST run after the+ # ``with`` block above has released the lock — nesting it inside would+ # self-deadlock the calling thread. Doing it here also means a slow+ # model load doesn't block concurrent ``set_llm_config()`` writes.return get_llm_backend_for_engine("qwen_llm")