[fix] load_audio: fall back to soundfile when torchcodec fails on WAV with trailing metadata (#24185)

python/sglang/srt/utils/common.py

@@ -782,17 +782,24 @@ def load_audio(
     if _BACKEND == "torchcodec":
         from torchcodec.decoders import AudioDecoder
 
-        decoder = AudioDecoder(
-            source,
-            sample_rate=sr,
-            num_channels=1 if mono else None,
-        )
-        samples = decoder.get_all_samples()
-        if mono:
-            return samples.data.squeeze(0).numpy()
-        return samples.data.T.numpy()
+        try:
+            decoder = AudioDecoder(
+                source,
+                sample_rate=sr,
+                num_channels=1 if mono else None,
+            )
+            samples = decoder.get_all_samples()
+            if mono:
+                return samples.data.squeeze(0).numpy()
+            return samples.data.T.numpy()
+        except Exception as e:
+            # torchcodec's bytes-buffer IO can fail on WAV files that carry
+            # large trailing metadata chunks. Fall back to soundfile, which reads the PCM payload directly.
+            logger.warning(
+                f"torchcodec AudioDecoder failed ({e}); falling back to soundfile + torchaudio."
+            )
 
-    # Fallback: soundfile + torchaudio (ARM / no FFmpeg)
+    # Fallback: soundfile + torchaudio (ARM / no FFmpeg / torchcodec failure)
     import soundfile as sf
     import torch
     import torchaudio