[Bugfix] Re-enable benchmarking of librispeech dataset. (#47033)

Signed-off-by: Anna Mayne <[email protected]>

tests/benchmarks/test_audio_dataset.py

@@ -123,15 +123,22 @@ def test_asr_dataset_sample_handles_local_audio_paths(tmp_path: Path) -> None:
     )
 
 
-def test_asr_dataset_sample_handles_embedded_audio_bytes(tmp_path: Path) -> None:
+@pytest.mark.parametrize("has_filepath", [True, False])
+def test_asr_dataset_sample_handles_embedded_audio_bytes(
+    tmp_path: Path, has_filepath: bool
+) -> None:
     audio_path = tmp_path / "earnings.wav"
     _write_wav(audio_path, duration_s=0.1)
 
+    test_path = None
+    if has_filepath:
+        test_path = audio_path
+
     dataset = object.__new__(datasets_module.ASRDataset)
     dataset.data = [
         {
             "audio": {
-                "path": None,
+                "path": test_path,
                 "bytes": audio_path.read_bytes(),
             },
             "text": "quarterly earnings call",

vllm/benchmarks/datasets/datasets.py

@@ -4090,9 +4090,10 @@ class ASRDataset(HuggingFaceDataset):
     EARNINGS22_TINY_FILTERED_DATASET = (
         "D4nt3/esb-datasets-earnings22-validation-tiny-filtered"
     )
+    LIBRISPEECH_DATASET = "openslr/librispeech_asr"
 
     SUPPORTED_DATASET_PATHS = {
-        "openslr/librispeech_asr",
+        LIBRISPEECH_DATASET,
         "facebook/voxpopuli",
         "LIUM/tedlium",
         "edinburghcstr/ami",
@@ -4120,7 +4121,10 @@ class ASRDataset(HuggingFaceDataset):
                 self.data = self.data.shuffle(seed=self.random_seed)
             self._materialize_local_audio_column()
             return
-        if self.hf_name == self.EARNINGS22_TINY_FILTERED_DATASET:
+        if self.hf_name in (
+            self.EARNINGS22_TINY_FILTERED_DATASET,
+            self.LIBRISPEECH_DATASET,
+        ):
             super().load_data()
             self._disable_audio_decode()
             return
@@ -4199,14 +4203,14 @@ class ASRDataset(HuggingFaceDataset):
             elif isinstance(audio, str):
                 duration_s = sf.info(audio).duration
                 mm_content = {"audio_path": audio}
-            elif isinstance(audio, dict) and audio.get("path"):
-                duration_s = sf.info(audio["path"]).duration
-                mm_content = {"audio_path": audio["path"]}
             elif isinstance(audio, dict) and audio.get("bytes") is not None:
                 with BytesIO(audio["bytes"]) as audio_buffer:
                     y, sr = sf.read(audio_buffer, dtype="float32")
                 duration_s = get_audio_duration(y=y, sr=sr)
                 mm_content = {"audio": (y, sr)}
+            elif isinstance(audio, dict) and audio.get("path"):
+                duration_s = sf.info(audio["path"]).duration
+                mm_content = {"audio_path": audio["path"]}
             else:
                 raise ValueError(
                     "ASR samples must provide decoded audio arrays, "