Move audio assets to conftest

Signed-off-by: Alex-Brooks <[email protected]>

Author: alex-jw-brooks

tests/conftest.py

@@ -21,6 +21,7 @@
 from tests.models.utils import (TokensTextLogprobs,
                                 TokensTextLogprobsPromptLogprobs)
 from vllm import LLM, SamplingParams
+from vllm.assets.audio import AudioAsset
 from vllm.assets.image import ImageAsset
 from vllm.assets.video import VideoAsset
 from vllm.config import TaskOption, _get_and_verify_dtype
@@ -103,10 +104,25 @@ def prompts(self, prompts: _VideoAssetPrompts) -> list[str]:
         return [prompts["sample_demo_1"]]
 
 
+class _AudioAssetsBase(UserList[AudioAsset]):
+    pass
+
+
+class _AudioAssets(_AudioAssetsBase):
+
+    def __init__(self) -> None:
+        super().__init__([
+            AudioAsset("mary_had_lamb"),
+            AudioAsset("winning_call"),
+        ])
+
+
 IMAGE_ASSETS = _ImageAssets()
 """Singleton instance of :class:`_ImageAssets`."""
 VIDEO_ASSETS = _VideoAssets()
 """Singleton instance of :class:`_VideoAssets`."""
+AUDIO_ASSETS = _AudioAssets()
+"""Singleton instance of :class:`_AudioAssets`."""
 
 
 @pytest.fixture(scope="function", autouse=True)
@@ -263,6 +279,11 @@ def video_assets() -> _VideoAssets:
     return VIDEO_ASSETS
 
 
+@pytest.fixture(scope="session")
+def audio_assets() -> _AudioAssets:
+    return AUDIO_ASSETS
+
+
 _T = TypeVar("_T", nn.Module, torch.Tensor, BatchEncoding, BatchFeature, dict)
 _R = TypeVar("_R")
 
@@ -390,10 +411,15 @@ def get_inputs(
                 processor_kwargs["images"] = image
             if videos is not None and (video := videos[i]) is not None:
                 processor_kwargs["videos"] = video
-            if audios is not None and (audio_tuple := audios[i]) is not None:
-                audio, sr = audio_tuple
-                processor_kwargs["audio"] = audio
-                processor_kwargs["sampling_rate"] = sr
+            if audios is not None and (audio_inputs := audios[i]) is not None:
+                # HACK - not all processors take sampling_rate; we should
+                # clean this up in the future.
+                if len(audio_inputs) == 2:
+                    audio, sr = audio_inputs
+                    processor_kwargs["audio"] = audio
+                    processor_kwargs["sampling_rate"] = sr
+                else:
+                    processor_kwargs["audio"] = audio_inputs
 
             inputs = self.processor(**processor_kwargs)
             if isinstance(inputs, BatchFeature):

tests/models/decoder_only/audio_language/test_ultravox.py

@@ -11,7 +11,7 @@
 from vllm.multimodal.audio import resample_audio_librosa
 from vllm.sequence import SampleLogprobs
 
-from ....conftest import HfRunner, VllmRunner
+from ....conftest import HfRunner, VllmRunner, _AudioAssets
 from ....utils import RemoteOpenAIServer
 from ...registry import HF_EXAMPLE_MODELS
 from ...utils import check_logprobs_close
@@ -31,12 +31,6 @@
 }
 
 
-@pytest.fixture(scope="session")
-def audio_assets():
-    from vllm.assets.audio import AudioAsset
-    return [AudioAsset("mary_had_lamb"), AudioAsset("winning_call")]
-
-
 @pytest.fixture(scope="module", params=("mary_had_lamb", "winning_call"))
 def audio(request):
     from vllm.assets.audio import AudioAsset
@@ -59,7 +53,7 @@ def params_kwargs_to_cli_args(params_kwargs: dict[str, Any]) -> list[str]:
     pytest.param({}, marks=pytest.mark.cpu_model),
     pytest.param(CHUNKED_PREFILL_KWARGS),
 ])
-def server(request, audio_assets):
+def server(request, audio_assets: _AudioAssets):
     args = [
         "--dtype", "bfloat16", "--max-model-len", "4096", "--enforce-eager",
         "--limit-mm-per-prompt",
@@ -230,8 +224,9 @@ def test_models(hf_runner, vllm_runner, audio, dtype: str, max_tokens: int,
     pytest.param({}, marks=pytest.mark.cpu_model),
     pytest.param(CHUNKED_PREFILL_KWARGS),
 ])
-def test_models_with_multiple_audios(vllm_runner, audio_assets, dtype: str,
-                                     max_tokens: int, num_logprobs: int,
+def test_models_with_multiple_audios(vllm_runner, audio_assets: _AudioAssets,
+                                     dtype: str, max_tokens: int,
+                                     num_logprobs: int,
                                      vllm_kwargs: dict) -> None:
 
     vllm_prompt = _get_prompt(len(audio_assets),
@@ -250,7 +245,7 @@ def test_models_with_multiple_audios(vllm_runner, audio_assets, dtype: str,
 
 
 @pytest.mark.asyncio
-async def test_online_serving(client, audio_assets):
+async def test_online_serving(client, audio_assets: _AudioAssets):
     """Exercises online serving with/without chunked prefill enabled."""
 
     messages = [{