feat: generate a book with all installed models to compare

2 files changed, 42 insertions, 8 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"",
diff --git a/app/converter/converter.py b/app/converter/converter.py
index bd5e477..a10a57d 100644
--- a/app/converter/converter.py
+++ b/app/converter/converter.py
@@ -212,6 +212,7 @@ class AudiobookConverter:
                  instructions: Optional[str] = None,
                  request_options: Optional[Dict[str, str]] = None,
                  api_url: Optional[str] = None,
+                 unload_models: Optional[bool] = None,
                  progress: Optional[Callable[[dict], None]] = None,
                  cancel=None):
         if speed <= 0:
@@ -243,6 +244,9 @@ class AudiobookConverter:
         # them against the server-hosted model at connect time.
         self.instructions = instructions
         self.request_options = dict(request_options or {})
+        # audio.cpp only: force unloading previously-loaded server models
+        # at connect time (None follows the AUDIOCPP_UNLOAD_MODELS setting).
+        self.unload_models = unload_models
         self._validate_configuration()
         # Interactive reporting (the TUI run view): PROGRESS receives an
         # event dict per state change and turns the clients' console prints
@@ -261,13 +265,15 @@ class AudiobookConverter:
             # elsewhere. model_id picks the server entry per run
             # (auto-selected on single-entry servers); instructions
             # describe or style the voice, request_options pass
-            # per-model controls through to the server.
+            # per-model controls through to the server. unload_models
+            # forces a pre-run model unload when not None ("All" runs).
             self.tts = AudioCppTTSClient(chunks_dir=CHUNKS_FOLDER,
                                          voice=voice, language=self.language,
                                          model_id=model_id,
                                          instructions=instructions,
                                          request_options=self.request_options,
-                                         api_url=api_url, quiet=quiet)
+                                         api_url=api_url, quiet=quiet,
+                                         unload_models=unload_models)
         else:
             # Qwen: the voice mode picks the request shape (built-in
             # speaker, clone from a reference .wav, or a designed voice);
@@ -355,7 +361,6 @@ class AudiobookConverter:
                              voice_clone_ref_audio: Optional[str],
                              instructions: Optional[str] = None) -> str:
         """Narrator name used in output file names, without a server connection.
-
         Custom voice mode uses the built-in speaker's display name; voice
         clone mode uses the reference audio file's stem; the faster and
         audiocpp backends use the server-side voice name (for audiocpp's
@@ -392,6 +397,19 @@ class AudiobookConverter:
         return AudiobookConverter._sanitize_filename(
             narrator, fallback="narrator").replace(" ", "_")
 
+    @staticmethod
+    def compute_model_tag(model_id: Optional[str]) -> str:
+        """Model id used in output file names, without a server connection.
+
+        "All (multiple generation)" runs name every output with the
+        generating model's id so the per-model files never collide
+        (e.g. ``dune_qwen3_tts_1_7b_base_q8_0_Vivian.m4b``). Pure (no I/O,
+        no server) so the pre-flight can compute the exact output names a
+        run would produce before spending time connecting to a TTS server.
+        """
+        return AudiobookConverter._sanitize_filename(
+            model_id or "", fallback="model").replace(" ", "_")
+
     # ------------------------------------------------------------------
     # Debug dumps (--debug)
     # ------------------------------------------------------------------
@@ -804,6 +822,7 @@ class AudiobookConverter:
                              confirm: Optional[Callable[[str, bool], bool]] = None,
                              book_files: Optional[List[Path]] = None,
                              output_name: Optional[str] = None,
+                             name_tag: Optional[str] = None,
                              ) -> Tuple[List[Path], List[Tuple[Path, str]]]:
         """Discover books and ask every overwrite question up front.
 
@@ -822,7 +841,10 @@ class AudiobookConverter:
         (a single --input-file book; still filtered to supported formats),
         and OUTPUT_NAME overrides the computed output name with a verbatim
         base name (--output-file's stem, no narrator tag or stem-collision
-        suffix). Both default to the directory-scan behavior.
+        suffix). Both default to the directory-scan behavior. NAME_TAG, when
+        given, is inserted between the book stem and the narrator tag
+        ("All (multiple generation)" runs pass the sanitized model id, so
+        each model's outputs are named and planned separately).
         """
         if book_files is None:
             book_files = sorted(
@@ -854,7 +876,8 @@ class AudiobookConverter:
                 name = book_file.stem
                 if stem_counts[book_file.stem] > 1:
                     name = f"{book_file.stem}_{book_file.suffix.lstrip('.')}"
-                names.append((book_file, f"{name}_{narrator_tag}"))
+                tag = f"{name_tag}_{narrator_tag}" if name_tag else narrator_tag
+                names.append((book_file, f"{name}_{tag}"))
 
         # Ask every overwrite question up front, before any conversion
         # starts, so the rest of the run is unattended.