feat: option to unload models from audio.cpp per-run

2 files changed, 11 insertions, 1 deletions

diff --git a/app/converter/config.py b/app/converter/config.py
index 9f70a73..6eacb41 100644
--- a/app/converter/config.py
+++ b/app/converter/config.py
@@ -83,3 +83,10 @@ AUDIOCPP_CLONE_MODEL_ID = "qwen"
 # VoiceDesign); on other families it acts as a style/delivery instruction
 # when the model supports one and is ignored otherwise. Empty by default.
 AUDIOCPP_INSTRUCTIONS = ""
+
+# Ask the audio.cpp server to unload all currently loaded models before
+# converting, so models left resident by earlier runs free their memory
+# (e.g. VRAM on GPU backends) and only the selected entry loads. Set to
+# False to keep other models resident across runs (the TUI Settings menu
+# exposes this as "Unload models", default Yes).
+AUDIOCPP_UNLOAD_MODELS = True
diff --git a/app/converter/tts.py b/app/converter/tts.py
index 41e3aa5..a83896d 100644
--- a/app/converter/tts.py
+++ b/app/converter/tts.py
@@ -919,7 +919,8 @@ class AudioCppTTSClient(_BaseTTSClient):
             print(f"[INFO] Sending instruction with every request: {self.instructions}")
             print("[INFO] Its effect (style, emotion, delivery) depends on the "
                   "model family; models without instruction support ignore it.")
-        self._unload_server_models()
+        if config.AUDIOCPP_UNLOAD_MODELS:
+            self._unload_server_models()
 
     def _unload_server_models(self) -> None:
         """Ask the server to unload every loaded model before generating.
@@ -930,6 +931,8 @@ class AudioCppTTSClient(_BaseTTSClient):
         first frees those leftovers; this run's model reloads transparently
         on its first request. Failures only warn: an older server without
         the endpoint, or a busy one, must not block a working setup.
+        Controlled by config.AUDIOCPP_UNLOAD_MODELS (the TUI Settings
+        "Unload models" option).
         """
         request = urllib.request.Request(
             f"{self.api_url}/v1/tasks/unload_all_models", data=b"",