|
| 1 | +import sys |
| 2 | +import types |
| 3 | +import importlib.machinery |
| 4 | +from types import SimpleNamespace |
| 5 | + |
| 6 | +import numpy as np |
| 7 | + |
| 8 | + |
| 9 | +def _install_qwen_tts_stub() -> None: |
| 10 | + try: |
| 11 | + __import__("qwen_tts") |
| 12 | + return |
| 13 | + except ImportError: |
| 14 | + pass |
| 15 | + |
| 16 | + class _FakeQwen3TTSModel: |
| 17 | + def __init__(self, *args, **kwargs): |
| 18 | + self.args = args |
| 19 | + self.kwargs = kwargs |
| 20 | + |
| 21 | + @classmethod |
| 22 | + def from_pretrained(cls, *args, **kwargs): |
| 23 | + return cls(*args, **kwargs) |
| 24 | + |
| 25 | + def generate_custom_voice(self, text, speaker=None, **kwargs): |
| 26 | + return [np.zeros(24_000, dtype=np.float32)], 24_000 |
| 27 | + |
| 28 | + def generate_voice_clone(self, text, **kwargs): |
| 29 | + return [np.zeros(24_000, dtype=np.float32)], 24_000 |
| 30 | + |
| 31 | + def generate_voice_design(self, text, language=None, instruct=None, **kwargs): |
| 32 | + return [np.zeros(24_000, dtype=np.float32)], 24_000 |
| 33 | + |
| 34 | + def create_voice_clone_prompt(self, ref_audio=None, ref_text=None, **kwargs): |
| 35 | + return { |
| 36 | + "ref_audio": ref_audio, |
| 37 | + "ref_text": ref_text, |
| 38 | + } |
| 39 | + |
| 40 | + qwen_tts_module = types.ModuleType("qwen_tts") |
| 41 | + qwen_tts_module.Qwen3TTSModel = _FakeQwen3TTSModel |
| 42 | + qwen_tts_module.__spec__ = importlib.machinery.ModuleSpec("qwen_tts", loader=None) |
| 43 | + sys.modules["qwen_tts"] = qwen_tts_module |
| 44 | + |
| 45 | + |
| 46 | +_install_qwen_tts_stub() |
| 47 | + |
| 48 | + |
| 49 | +def _install_torchcodec_stub() -> None: |
| 50 | + try: |
| 51 | + __import__("torchcodec") |
| 52 | + return |
| 53 | + except ImportError: |
| 54 | + pass |
| 55 | + |
| 56 | + class _FakeTensor: |
| 57 | + def __init__(self, array): |
| 58 | + self._array = array |
| 59 | + |
| 60 | + def cpu(self): |
| 61 | + return self |
| 62 | + |
| 63 | + def numpy(self): |
| 64 | + return self._array |
| 65 | + |
| 66 | + class _FakeAudioSamples: |
| 67 | + def __init__(self, data=None, sample_rate: int = 16_000): |
| 68 | + _arr = np.zeros((1, sample_rate), dtype=np.float32) if data is None else data |
| 69 | + self.data = _FakeTensor(_arr) |
| 70 | + self.sample_rate = sample_rate |
| 71 | + |
| 72 | + class _FakeAudioDecoder: |
| 73 | + def __init__(self, source=None, *args, **kwargs): |
| 74 | + self.source = source |
| 75 | + self.args = args |
| 76 | + self.kwargs = kwargs |
| 77 | + _path = None |
| 78 | + if isinstance(source, dict): |
| 79 | + _path = source.get("path") |
| 80 | + else: |
| 81 | + _path = getattr(source, "path", None) |
| 82 | + |
| 83 | + self.metadata = SimpleNamespace( |
| 84 | + sample_rate=16_000, |
| 85 | + path=_path, |
| 86 | + ) |
| 87 | + |
| 88 | + def __getitem__(self, key: str): |
| 89 | + if key == "path": |
| 90 | + return self.metadata.path |
| 91 | + if key == "sampling_rate": |
| 92 | + return self.metadata.sample_rate |
| 93 | + if key == "array": |
| 94 | + y = self.get_all_samples().data.cpu().numpy() |
| 95 | + return np.mean(y, axis=tuple(range(y.ndim - 1))) if y.ndim > 1 else y |
| 96 | + raise KeyError(key) |
| 97 | + |
| 98 | + def get_all_samples(self): |
| 99 | + return _FakeAudioSamples() |
| 100 | + |
| 101 | + def get_samples_played_in_range(self, *_args, **_kwargs): |
| 102 | + return SimpleNamespace(sample_rate=self.metadata.sample_rate) |
| 103 | + |
| 104 | + torchcodec_module = types.ModuleType("torchcodec") |
| 105 | + decoders_module = types.ModuleType("torchcodec.decoders") |
| 106 | + decoders_module.AudioDecoder = _FakeAudioDecoder |
| 107 | + torchcodec_module.decoders = decoders_module |
| 108 | + torchcodec_module.__spec__ = importlib.machinery.ModuleSpec("torchcodec", loader=None) |
| 109 | + decoders_module.__spec__ = importlib.machinery.ModuleSpec("torchcodec.decoders", loader=None) |
| 110 | + |
| 111 | + sys.modules["torchcodec"] = torchcodec_module |
| 112 | + sys.modules["torchcodec.decoders"] = decoders_module |
| 113 | + |
| 114 | + # transformers.audio_utils (and others) call importlib.metadata.version("torchcodec") |
| 115 | + # at module level. That call bypasses sys.modules and reads on-disk dist-info, so |
| 116 | + # it raises PackageNotFoundError even though our stub is in sys.modules, causing a |
| 117 | + # cascade failure across ALL tests. Patch it to return a harmless version string. |
| 118 | + import importlib.metadata as _imeta |
| 119 | + _real_version = _imeta.version |
| 120 | + |
| 121 | + def _patched_version(name: str) -> str: |
| 122 | + if name == "torchcodec": |
| 123 | + return "0.0.0" |
| 124 | + return _real_version(name) |
| 125 | + |
| 126 | + _imeta.version = _patched_version |
| 127 | + |
| 128 | + |
| 129 | +_install_torchcodec_stub() |
0 commit comments