feat: generate a book with all installed models to compare

1 files changed, 14 insertions, 3 deletions

diff --git a/app/converter/clients/audiocpp.py b/app/converter/clients/audiocpp.py
index a1888bc..df1884d 100644
--- a/app/converter/clients/audiocpp.py
+++ b/app/converter/clients/audiocpp.py
@@ -406,7 +406,8 @@ class AudioCppTTSClient(BaseTTSClient):
                  model_id: Optional[str] = None,
                  instructions: Optional[str] = None,
                  request_options: Optional[Dict[str, str]] = None,
-                 quiet: bool = False):
+                 quiet: bool = False,
+                 unload_models: Optional[bool] = None):
         super().__init__(chunks_dir, quiet=quiet)
         self.api_url = (api_url or config.AUDIOCPP_API_URL).rstrip("/")
         # Per-run model selection: the --model CLI flag (or the Generate
@@ -415,6 +416,12 @@ class AudioCppTTSClient(BaseTTSClient):
         # don't require --model.
         self.model_id = (model_id or "").strip()
         self._model_id_explicit = bool(self.model_id)
+        # Unload previously-loaded server models at connect time: None
+        # follows the AUDIOCPP_UNLOAD_MODELS setting (read at connect, so
+        # a Settings change this session is honored); True/False force it
+        # regardless of the setting ("All (multiple generation)" runs pass
+        # True so each per-model conversion starts with a clean VRAM).
+        self._unload_models_override = unload_models
         # Validate before connecting so bad values fail fast without a server.
         self.language = normalize_language(
             language if language is not None else config.LANGUAGE)
@@ -570,7 +577,10 @@ class AudioCppTTSClient(BaseTTSClient):
             self._report(f"[INFO] Sending instruction with every request: {self.instructions}")
             self._report("[INFO] Its effect (style, emotion, delivery) depends on the "
                          "model family; models without instruction support ignore it.")
-        if config.AUDIOCPP_UNLOAD_MODELS:
+        unload = (config.AUDIOCPP_UNLOAD_MODELS
+                  if self._unload_models_override is None
+                  else self._unload_models_override)
+        if unload:
             self._unload_server_models()
 
     def _require_synthesis_task(self, models: List[Dict[str, str]]) -> None:
@@ -595,7 +605,8 @@ class AudioCppTTSClient(BaseTTSClient):
         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).
+        "Unload models" option), or forced per run via the unload_models
+        override ("All (multiple generation)" runs unload between models).
         """
         request = urllib.request.Request(
             f"{self.api_url}/v1/tasks/unload_all_models", data=b"",