From 919544c0931d53bb81904b6212ff14f856549da3 Mon Sep 17 00:00:00 2001 From: historia Date: Mon, 24 Aug 2026 16:18:38 -0400 Subject: feat: option to unload models from audio.cpp per-run --- app/converter/config.py | 7 +++++++ app/converter/tts.py | 5 ++++- app/docs/backend-audiocpp.md | 2 +- app/tests/test_hub.py | 26 +++++++++++++++++++------- app/tests/test_tts.py | 30 ++++++++++++++++++++++++++++++ app/ui/hub.py | 5 +++++ 6 files changed, 66 insertions(+), 9 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"", diff --git a/app/docs/backend-audiocpp.md b/app/docs/backend-audiocpp.md index aaeb5a5..5ce7f49 100644 --- a/app/docs/backend-audiocpp.md +++ b/app/docs/backend-audiocpp.md @@ -92,4 +92,4 @@ python audiobook.py --backend audiocpp --model qwen-design \ The hub also works with an audio.cpp server that runs somewhere else (another checkout, another machine): set `AUDIOCPP_REMOTE_URL` in `app/converter/config.py` (or the TUI **Settings** → "audio.cpp remote URL") to its `host:port`. The hub probes that URL and, when it answers, offers an `audio.cpp [remote]` entry in **Convert books…** whose models and voices are queried live (`GET /v1/models` and `GET /v1/audio/voices`) — alongside the managed `audio.cpp` entry, which keeps reading the local `server.json`. The remote URL defaults to `127.0.0.1:8080`, so a server started outside this tool on the local port is found automatically. On the CLI, pass `--api-url http://host:port` (and `--model`/`--voice` matching that server's config). -Before converting, `audiobook.py` asks the server to unload all currently loaded models (`POST /v1/tasks/unload_all_models`) so models left resident by earlier runs free their memory (e.g. VRAM on GPU backends) and only the selected entry loads. A server without that endpoint, or one busy unloading, only produces a warning. +Before converting, `audiobook.py` asks the server to unload all currently loaded models (`POST /v1/tasks/unload_all_models`) so models left resident by earlier runs free their memory (e.g. VRAM on GPU backends) and only the selected entry loads. A server without that endpoint, or one busy unloading, only produces a warning. This behavior is controlled by the **Settings** → "Unload models" option (or `AUDIOCPP_UNLOAD_MODELS` in `app/converter/config.py`), which defaults to **Yes**; set it to **No** to keep other models resident across runs. diff --git a/app/tests/test_hub.py b/app/tests/test_hub.py index 725f4f6..a0545cc 100644 --- a/app/tests/test_hub.py +++ b/app/tests/test_hub.py @@ -1083,7 +1083,8 @@ class SettingsTests(unittest.TestCase): original = {name: getattr(hub.config, name) for name in ("AUDIO_FORMAT", "AUDIO_BITRATE", "LANGUAGE", - "CHUNK_SIZE", "QWEN_API_URL", "CLONE_API_URL", + "CHUNK_SIZE", "AUDIOCPP_UNLOAD_MODELS", + "QWEN_API_URL", "CLONE_API_URL", "FASTER_API_URL", "AUDIOCPP_API_URL", "QWEN_REMOTE_URL", "CLONE_REMOTE_URL", "FASTER_REMOTE_URL", "AUDIOCPP_REMOTE_URL")} @@ -1091,6 +1092,7 @@ class SettingsTests(unittest.TestCase): for name, value in original.items()]) values = {"audio_format": "ogg", "audio_bitrate": " 192k ", "language": "en", "chunk_size": "300", + "unload_models": True, "qwen_custom_port": "7862", "qwen_clone_port": "7863", "faster_port": "8001", "audiocpp_port": "8081", "audiocpp_remote_url": "10.0.0.5:8080", @@ -1106,6 +1108,7 @@ class SettingsTests(unittest.TestCase): "AUDIO_BITRATE": "192k", "LANGUAGE": "English", "CHUNK_SIZE": 300, + "AUDIOCPP_UNLOAD_MODELS": True, "QWEN_API_URL": "http://127.0.0.1:7862", "CLONE_API_URL": "http://127.0.0.1:7863", "FASTER_API_URL": "http://127.0.0.1:8001", @@ -1122,6 +1125,7 @@ class SettingsTests(unittest.TestCase): self.assertEqual(hub.config.AUDIO_BITRATE, "192k") self.assertEqual(hub.config.LANGUAGE, "English") self.assertEqual(hub.config.CHUNK_SIZE, 300) + self.assertEqual(hub.config.AUDIOCPP_UNLOAD_MODELS, True) self.assertEqual(hub.config.QWEN_API_URL, "http://127.0.0.1:7862") self.assertEqual(hub.config.FASTER_API_URL, "http://127.0.0.1:8001") self.assertEqual(hub.config.AUDIOCPP_REMOTE_URL, @@ -1130,7 +1134,8 @@ class SettingsTests(unittest.TestCase): def test_apply_settings_rejects_bad_values(self): original = {name: getattr(hub.config, name) for name in ("AUDIO_FORMAT", "AUDIO_BITRATE", "LANGUAGE", - "CHUNK_SIZE", "QWEN_API_URL", "CLONE_API_URL", + "CHUNK_SIZE", "AUDIOCPP_UNLOAD_MODELS", + "QWEN_API_URL", "CLONE_API_URL", "FASTER_API_URL", "AUDIOCPP_API_URL", "QWEN_REMOTE_URL", "CLONE_REMOTE_URL", "FASTER_REMOTE_URL", "AUDIOCPP_REMOTE_URL")} @@ -1138,6 +1143,7 @@ class SettingsTests(unittest.TestCase): for name, value in original.items()]) base = {"audio_format": "m4b", "audio_bitrate": "128k", "language": "English", "chunk_size": "250", + "unload_models": True, "qwen_custom_port": "7860", "qwen_clone_port": "7861", "faster_port": "8000", "audiocpp_port": "8080"} with patch.object(hub, "_write_config") as mk_write: @@ -1175,6 +1181,7 @@ class SettingsTests(unittest.TestCase): captured["fields"] = fields return {"audio_format": "ogg", "audio_bitrate": "192k", "language": "English", "chunk_size": "300", + "unload_models": True, "qwen_custom_port": "7860", "qwen_clone_port": "7861", "faster_port": "8000", "audiocpp_port": "8080"} @@ -1192,14 +1199,16 @@ class SettingsTests(unittest.TestCase): hub._settings_menu(None) self.assertEqual([f["key"] for f in captured["fields"]], ["audio_format", "audio_bitrate", "language", - "chunk_size", "audiocpp_port", "faster_port", - "qwen_custom_port", "qwen_clone_port", - "audiocpp_remote_url", "faster_remote_url", - "qwen_custom_remote_url", "qwen_clone_remote_url"]) + "chunk_size", "unload_models", "audiocpp_port", + "faster_port", "qwen_custom_port", + "qwen_clone_port", "audiocpp_remote_url", + "faster_remote_url", "qwen_custom_remote_url", + "qwen_clone_remote_url"]) kinds = {f["key"]: f["kind"] for f in captured["fields"]} self.assertEqual(kinds["audio_format"], "choice") self.assertEqual(kinds["audio_bitrate"], "text") self.assertEqual(kinds["audiocpp_port"], "text") + self.assertEqual(kinds["unload_models"], "bool") self.assertEqual(kinds["audiocpp_remote_url"], "text") labels = {f["key"]: f["label"] for f in captured["fields"]} self.assertEqual(labels["qwen_clone_port"], "qwen-tts Base port") @@ -1217,6 +1226,7 @@ class SettingsTests(unittest.TestCase): "audio_bitrate": "192k", "language": "English", "chunk_size": "300", + "unload_models": True, "qwen_custom_port": "7860", "qwen_clone_port": "7861", "faster_port": "8000", @@ -1248,7 +1258,8 @@ class SettingsTests(unittest.TestCase): original = {name: getattr(hub.config, name) for name in ("AUDIO_FORMAT", "AUDIO_BITRATE", "LANGUAGE", - "CHUNK_SIZE", "QWEN_API_URL", "CLONE_API_URL", + "CHUNK_SIZE", "AUDIOCPP_UNLOAD_MODELS", + "QWEN_API_URL", "CLONE_API_URL", "FASTER_API_URL", "AUDIOCPP_API_URL", "QWEN_REMOTE_URL", "CLONE_REMOTE_URL", "FASTER_REMOTE_URL", "AUDIOCPP_REMOTE_URL")} @@ -1264,6 +1275,7 @@ class SettingsTests(unittest.TestCase): 'LANGUAGE = "English"\n' "\n" "CHUNK_SIZE = 250\n" + "AUDIOCPP_UNLOAD_MODELS = True\n" 'QWEN_API_URL = "http://127.0.0.1:7860"\n' 'CLONE_API_URL = "http://127.0.0.1:7861"\n' 'FASTER_API_URL = "http://127.0.0.1:8000"\n' diff --git a/app/tests/test_tts.py b/app/tests/test_tts.py index 0026702..77f0ee8 100644 --- a/app/tests/test_tts.py +++ b/app/tests/test_tts.py @@ -1394,6 +1394,36 @@ class AudioCppUnloadModelsTests(unittest.TestCase): client._connect() mock_unload.assert_called_once() + def test_connect_skips_unload_when_disabled(self): + client = AudioCppTTSClient.__new__(AudioCppTTSClient) + client.api_url = "http://127.0.0.1:8080" + client.model_id = config.AUDIOCPP_MODEL_ID + client.preset_mode = True + client.voice = "narrator" + client.language = "English" + client._seed = -1 + client.family = "qwen3_tts" + client.task = tts.AUDIOCPP_TASK_TTS + client.profile = tts.AUDIOCPP_FAMILY_PROFILES["qwen3_tts"] + client.design_mode = False + client.instruction_voice = False + client.instructions = "" + with patch.object(client, "_check_health"), \ + patch.object(client, "_list_models", + return_value=[{"id": client.model_id, + "family": "qwen3_tts", + "task": "tts"}]), \ + patch.object(client, "_auto_pick_model_id"), \ + patch.object(client, "_select_model"), \ + patch.object(client, "_require_model_id"), \ + patch.object(client, "_resolve_family"), \ + patch.object(client, "_resolve_task"), \ + patch.object(client, "_check_voice"), \ + patch.object(tts.config, "AUDIOCPP_UNLOAD_MODELS", False), \ + patch.object(client, "_unload_server_models") as mock_unload: + client._connect() + mock_unload.assert_not_called() + class BackendWiringTests(unittest.TestCase): """AudiobookConverter wiring for the --backend selector.""" diff --git a/app/ui/hub.py b/app/ui/hub.py index 632ff93..77d0796 100644 --- a/app/ui/hub.py +++ b/app/ui/hub.py @@ -661,6 +661,10 @@ def _settings_menu(stdscr) -> None: "value": config.LANGUAGE, "validate": _validate_language}, {"key": "chunk_size", "label": "Chunk size (words)", "kind": "text", "value": str(config.CHUNK_SIZE), "validate": _validate_chunk_size}, + {"key": "unload_models", "label": "Unload models", "kind": "bool", + "value": config.AUDIOCPP_UNLOAD_MODELS, + "note": "audio.cpp only. Ask the server to unload resident models " + "before converting."}, {"key": "audiocpp_port", "label": "audio.cpp port", "kind": "text", "value": str(_port_from_url(config.AUDIOCPP_API_URL, 8080)), @@ -796,6 +800,7 @@ def _apply_settings(values: dict) -> None: "AUDIO_BITRATE": bitrate, "LANGUAGE": normalize_language(values["language"]), "CHUNK_SIZE": chunk_size, + "AUDIOCPP_UNLOAD_MODELS": bool(values["unload_models"]), "QWEN_API_URL": common.url_with_port( config.QWEN_API_URL, ports["qwen_custom_port"]), "CLONE_API_URL": common.url_with_port( -- cgit v1.2.3