| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for ``MistralCommonFeatureExtractor.fetch_audio``. |
| 4 | |
| 5 | ``transformers>=5.10`` adds a ``ProcessorMixin.prepare_inputs_layout`` helper |
| 6 | that calls ``self.feature_extractor.fetch_audio(...)`` unconditionally. The |
| 7 | duck-typed :class:`MistralCommonFeatureExtractor` previously did not implement |
| 8 | that method, so loading any voxtral model under transformers 5.10.x raised |
| 9 | ``AttributeError: 'MistralCommonFeatureExtractor' object has no attribute |
| 10 | 'fetch_audio'``. These tests pin the new ``fetch_audio`` method to the same |
| 11 | contract as ``transformers.SequenceFeatureExtractor.fetch_audio``. |
| 12 | """ |
| 13 | |
| 14 | import numpy as np |
| 15 | import pytest |
| 16 | import torch |
| 17 | |
| 18 | from vllm.tokenizers.mistral import MistralTokenizer |
| 19 | from vllm.transformers_utils.processors.voxtral import ( |
| 20 | MistralCommonFeatureExtractor, |
| 21 | ) |
| 22 | |
| 23 | |
| 24 | @pytest.fixture(scope="module") |
| 25 | def feature_extractor() -> MistralCommonFeatureExtractor: |
| 26 | tokenizer = MistralTokenizer.from_pretrained("mistralai/Voxtral-Mini-3B-2507") |
| 27 | return MistralCommonFeatureExtractor(tokenizer.instruct.audio_encoder) |
| 28 | |
| 29 | |
| 30 | @pytest.mark.parametrize( |
| 31 | "audio", |
| 32 | [ |
| 33 | np.zeros(1024, dtype=np.float32), |
| 34 | torch.zeros(1024), |
| 35 | [0.0, 1.0, 2.0], |
| 36 | ], |
| 37 | ids=["numpy_array", "torch_tensor", "list_of_floats"], |
| 38 | ) |
| 39 | def test_fetch_audio_passes_through( |
| 40 | feature_extractor: MistralCommonFeatureExtractor, audio |
| 41 | ): |
| 42 | result = feature_extractor.fetch_audio(audio) |
| 43 | assert result is audio |
| 44 | |
| 45 | |
| 46 | def test_fetch_audio_recurses_over_list_of_arrays( |
| 47 | feature_extractor: MistralCommonFeatureExtractor, |
| 48 | ): |
| 49 | a = np.zeros(8, dtype=np.float32) |
| 50 | b = np.ones(8, dtype=np.float32) |
| 51 | result = feature_extractor.fetch_audio([a, b]) |
| 52 | assert isinstance(result, list) |
| 53 | assert len(result) == 2 |
| 54 | assert result[0] is a |
| 55 | assert result[1] is b |
| 56 | |
| 57 | |
| 58 | def test_fetch_audio_uses_self_sampling_rate_when_none( |
| 59 | monkeypatch, feature_extractor: MistralCommonFeatureExtractor |
| 60 | ): |
| 61 | """If ``sampling_rate`` is None, ``self.sampling_rate`` must be used. |
| 62 | |
| 63 | Verified indirectly via the recursion path: when we pass a list of arrays |
| 64 | without sampling_rate, recursive calls receive the resolved rate. |
| 65 | """ |
| 66 | captured: list[int | None] = [] |
| 67 | original = feature_extractor.fetch_audio |
| 68 | |
| 69 | def spy(audio, sampling_rate=None): |
| 70 | captured.append(sampling_rate) |
| 71 | return original(audio, sampling_rate=sampling_rate) |
| 72 | |
| 73 | monkeypatch.setattr(feature_extractor, "fetch_audio", spy) |
| 74 | feature_extractor.fetch_audio([np.zeros(4, dtype=np.float32)]) |
| 75 | # Top-level call has sampling_rate=None; inner recursive call sees the |
| 76 | # resolved rate from self.sampling_rate. |
| 77 | assert captured[0] is None |
| 78 | assert captured[1] == 16000 |
| 79 | |
| 80 | |
| 81 | def test_fetch_audio_explicit_sampling_rate_propagates( |
| 82 | monkeypatch, feature_extractor: MistralCommonFeatureExtractor |
| 83 | ): |
| 84 | captured: list[int | None] = [] |
| 85 | original = feature_extractor.fetch_audio |
| 86 | |
| 87 | def spy(audio, sampling_rate=None): |
| 88 | captured.append(sampling_rate) |
| 89 | return original(audio, sampling_rate=sampling_rate) |
| 90 | |
| 91 | monkeypatch.setattr(feature_extractor, "fetch_audio", spy) |
| 92 | feature_extractor.fetch_audio([np.zeros(4, dtype=np.float32)], sampling_rate=8000) |
| 93 | assert captured[0] == 8000 |
| 94 | assert captured[1] == 8000 |
| 95 | |
| 96 | |
| 97 | def test_fetch_audio_rejects_unsupported_type( |
| 98 | feature_extractor: MistralCommonFeatureExtractor, |
| 99 | ): |
| 100 | with pytest.raises(TypeError, match="only a numpy array"): |
| 101 | feature_extractor.fetch_audio(42) |
| 102 | |
| 103 | |
| 104 | def test_fetch_audio_str_delegates_to_load_audio( |
| 105 | monkeypatch, feature_extractor: MistralCommonFeatureExtractor |
| 106 | ): |
| 107 | """A str input must round-trip through ``transformers.audio_utils.load_audio``. |
| 108 | |
| 109 | We monkey-patch ``load_audio`` so the test stays offline (no real URL/path |
| 110 | fetched) and still asserts the delegation contract. |
| 111 | """ |
| 112 | sentinel = np.array([0.5, -0.5], dtype=np.float32) |
| 113 | received: dict[str, object] = {} |
| 114 | |
| 115 | def fake_load_audio(path, sampling_rate=None): |
| 116 | received["path"] = path |
| 117 | received["sampling_rate"] = sampling_rate |
| 118 | return sentinel |
| 119 | |
| 120 | import transformers.audio_utils |
| 121 | |
| 122 | monkeypatch.setattr(transformers.audio_utils, "load_audio", fake_load_audio) |
| 123 | |
| 124 | result = feature_extractor.fetch_audio("/tmp/fake.wav") |
| 125 | assert result is sentinel |
| 126 | assert received["path"] == "/tmp/fake.wav" |
| 127 | assert received["sampling_rate"] == 16000 |