[Bugfix][Audio] Restore soundfile-first automatic decoding (#55642)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>

docs/features/multimodal_inputs.md

@@ -1210,7 +1210,7 @@ vLLM decodes audio bytes into waveforms using a selectable decoding backend:
 
 | Backend | Description |
 | --- | --- |
-| `auto` (default) | torchcodec, falling back to soundfile, then PyAV |
+| `auto` (default) | soundfile, falling back to torchcodec, then PyAV |
 | `soundfile` | libsndfile only, no fallback |
 | `pyav` | PyAV (FFmpeg) only, no fallback |
 | `torchcodec` | TorchCodec (PyTorch-native) only, no fallback |
@@ -1226,15 +1226,17 @@ vllm serve mistralai/Voxtral-Mini-3B-2507 \
     `pyav` drives FFmpeg through a per-frame Python generator, so under
     concurrency the Python/C crossings contend on the GIL. `torchcodec`
     decodes each stream in a single call that releases the GIL for its whole
-    duration, which is why `auto` prefers it when many requests decode audio
-    concurrently, such as when audio tracks are extracted from video.
+    duration. Select it explicitly for concurrent decoding workloads.
+    `auto` prefers soundfile for supported formats to preserve their existing
+    decoding behavior, including encoder padding. Audio extracted from formats
+    that soundfile cannot read, such as video containers, can use torchcodec
+    through the fallback chain.
 
 !!! note
     `torchcodec` ships as a requirement on CUDA, CPU and XPU builds. On other
-    platforms (e.g. ROCm, TPU) `auto` falls back to the soundfile → PyAV
-    chain unless you install it manually. torchcodec also links against a
-    system FFmpeg installation; if FFmpeg is missing, `auto` falls back the
-    same way.
+    platforms (e.g. ROCm, TPU), install it manually to enable that backend.
+    torchcodec also links against a system FFmpeg installation. If the package
+    or FFmpeg is unavailable, `auto` uses the soundfile → PyAV chain.
 
 ### Embedding Inputs
 

tests/multimodal/media/test_audio.py

@@ -193,6 +193,14 @@ def test_load_audio_backend_matches_default(backend, dummy_audio_bytes):
     np.testing.assert_allclose(ref_audio[:n], audio[:n], atol=1e-4)
 
 
+def test_load_audio_default_preserves_vorbis_length(dummy_audio_bytes):
+    """Default decoding must not append Vorbis padding to the speech waveform."""
+    expected, expected_sr = load_audio_soundfile(BytesIO(dummy_audio_bytes), sr=None)
+    audio, sr = load_audio(BytesIO(dummy_audio_bytes), sr=None)
+    assert sr == expected_sr
+    assert audio.shape == expected.shape
+
+
 def test_load_audio_unknown_backend_rejected(dummy_audio_bytes):
     """An unknown backend must fail loudly instead of silently degrading."""
     with pytest.raises(ValueError, match="Unknown audio backend"):

vllm/multimodal/media/audio.py

@@ -60,11 +60,11 @@ _BAD_SF_CODES = {0, 1, 3, 4}
 
 # Audio decoding backends selectable via `load_audio(backend=...)` or
 # `--media-io-kwargs '{"audio": {"audio_backend": ...}}'`.
-# "auto" tries torchcodec, then soundfile, then PyAV.
+# "auto" tries soundfile, then torchcodec, then PyAV.
 AUDIO_BACKENDS = ("auto", "soundfile", "pyav", "torchcodec")
 
 # Raised as ImportError when the torchcodec backend cannot be used, so
-# `load_audio(backend="auto")` falls back to the soundfile → PyAV chain.
+# `load_audio(backend="auto")` can continue to PyAV.
 _TORCHCODEC_UNAVAILABLE_MSG = (
     "torchcodec audio backend is unavailable (requires the torchcodec "
     "package and a system ffmpeg installation)"
@@ -413,8 +413,8 @@ def load_audio(
 
     Args:
         backend: One of ``AUDIO_BACKENDS``. ``None`` (default) selects
-            ``"auto"``, which tries torchcodec, then falls back to the
-            soundfile → PyAV chain; the other values select a single
+            ``"auto"``, which tries soundfile, then torchcodec, then PyAV;
+            the other values select a single
             backend with no fallback.
     """
     backend = backend or "auto"
@@ -433,27 +433,11 @@ def load_audio(
             max_decode_bytes=max_decode_bytes,
         )
 
-    # "auto": torchcodec → soundfile → PyAV. Each backend is called by its
+    # Keep soundfile first to preserve decoding and padding of supported formats.
+    # "auto": soundfile → torchcodec → PyAV. Each backend is called by its
     # bare name so tests that monkeypatch it take effect. Only an ImportError
     # (backend missing) or, for soundfile, an unsupported-format error defers
     # to the next; any other error (decode failure, guard ValueError) raises.
-    if isinstance(path, BytesIO):
-        path.seek(0)
-    try:
-        return load_audio_torchcodec(
-            path,
-            sr=sr,
-            mono=mono,
-            max_duration_s=max_duration_s,
-            max_decode_bytes=max_decode_bytes,
-        )
-    except ImportError as exc:
-        # Decode errors don't defer: torchcodec is FFmpeg-based like PyAV, so
-        # retrying would just fail the same way.
-        logger.warning(
-            "torchcodec unavailable (%r); falling back to soundfile/PyAV.", exc
-        )
-
     if isinstance(path, BytesIO):
         path.seek(0)
     try:
@@ -477,6 +461,21 @@ def load_audio(
         if exc.code not in _BAD_SF_CODES:
             raise
 
+    if isinstance(path, BytesIO):
+        path.seek(0)
+    try:
+        return load_audio_torchcodec(
+            path,
+            sr=sr,
+            mono=mono,
+            max_duration_s=max_duration_s,
+            max_decode_bytes=max_decode_bytes,
+        )
+    except ImportError as exc:
+        # Decode errors don't defer: torchcodec is FFmpeg-based like PyAV, so
+        # retrying would just fail the same way.
+        logger.warning("torchcodec unavailable (%r); falling back to PyAV.", exc)
+
     # PyAV is terminal: nothing left to fall back to. Normalize an FFmpeg
     # failure to ValueError; let a missing soundfile/PyAV surface its
     # "install vllm[audio]" ImportError.