Write WAV via soundfile (torchaudio 2.10 delegates save to torchcodec)

app.py

@@ -57,6 +57,26 @@ logger.info("Release assets ready in %.1fs at %s", time.time() - _t0, RELEASE_RO
 os.environ["OMNIVAE_RELEASE_ROOT"] = RELEASE_ROOT
 CHECKPOINT_DIR = os.path.join(RELEASE_ROOT, "models", "dit", "t2av", EXPERIMENT)
 
+import torchaudio  # noqa: E402
+
+try:  # pragma: no cover
+    import torchcodec  # noqa: F401
+except ImportError:
+    # torchaudio >= 2.10 delegates `save` to torchcodec, which is not installed
+    # (its wheels are pinned to a specific FFmpeg ABI). The pipeline only needs
+    # a plain 48 kHz WAV write, so route it through soundfile instead.
+    import numpy as _np  # noqa: E402
+    import soundfile as _sf  # noqa: E402
+
+    def _save_with_soundfile(uri, src, sample_rate, **_kwargs):
+        data = src.detach().to("cpu", torch.float32).numpy()
+        if data.ndim == 2:  # torchaudio is (channels, time); soundfile wants (time, channels)
+            data = data.T
+        _sf.write(str(uri), _np.ascontiguousarray(data), int(sample_rate))
+
+    torchaudio.save = _save_with_soundfile
+    logger.info("torchcodec unavailable; torchaudio.save patched to use soundfile")
+
 from t2av_pipeline import generate_one_av, load_joint_av_pipeline  # noqa: E402
 
 # ---------------------------------------------------------------------------