Pin torch/torchvision/torchaudio to 2.2.2 trio for pyannote 3.4.0 compat (reverts the AudioMetaData shim, which didn't find the class)

app.py

@@ -156,34 +156,6 @@ def _get_diar_pipe():
     if _diar_pipe is not None:
         return _diar_pipe
     import os
-    import torchaudio
-
-    # Compat shim: pyannote.audio 3.4.0 expects `torchaudio.AudioMetaData` at
-    # the top level (true in torchaudio 2.2.x). Newer torchaudio releases
-    # dropped that re-export even though the underlying class still exists.
-    # Reattach it from wherever it actually lives, without touching the
-    # installed torch/torchaudio versions (which the image/video generation
-    # tabs on this Space depend on).
-    if not hasattr(torchaudio, "AudioMetaData"):
-        _meta_cls = None
-        for _modpath in (
-            "torchaudio._backend.common",
-            "torchaudio.backend.common",
-            "torchaudio.io",
-        ):
-            try:
-                _mod = __import__(_modpath, fromlist=["AudioMetaData"])
-                _meta_cls = getattr(_mod, "AudioMetaData", None)
-                if _meta_cls is not None:
-                    break
-            except ImportError:
-                continue
-        if _meta_cls is not None:
-            torchaudio.AudioMetaData = _meta_cls
-            print(f"Shimmed torchaudio.AudioMetaData from {_modpath}", flush=True)
-        else:
-            print("WARN: could not find AudioMetaData to shim onto torchaudio", flush=True)
-
     from pyannote.audio import Pipeline as DiarPipeline
 
     token = os.environ.get("HF_TOKEN")

requirements.txt

@@ -6,8 +6,9 @@ Pillow>=10.0
 scipy>=1.11
 imageio>=2.34
 imageio-ffmpeg>=0.5
-torch>=2.2
-torchvision>=0.17
+torch==2.2.2
+torchvision==0.17.2
+torchaudio==2.2.2
 diffusers>=0.38.0
 transformers>=4.40
 accelerate>=0.30