Revert torch pin (broke spaces/ZeroGPU + Wan2.2/LTX-Video); make diarization best-effort and drop pyannote.audio dep. Transcription-only stays GPU-accelerated.

app.py

@@ -200,13 +200,23 @@ def infer_audio(audio_path: str) -> tuple[str, str]:
             lines.append(f"[{start:07.2f} -> {end:07.2f}] {text}")
     transcript_text = "\n".join(lines) if lines else asr_result.get("text", "")
 
-    diar = _get_diar_pipe()
-    diarization = diar(audio_path)
-    turns = [
-        {"start": round(turn.start, 2), "end": round(turn.end, 2), "speaker": speaker}
-        for turn, _, speaker in diarization.itertracks(yield_label=True)
-    ]
-    diarization_json = json.dumps(turns, indent=2)
+    try:
+        diar = _get_diar_pipe()
+        diarization = diar(audio_path)
+        turns = [
+            {"start": round(turn.start, 2), "end": round(turn.end, 2), "speaker": speaker}
+            for turn, _, speaker in diarization.itertracks(yield_label=True)
+        ]
+        diarization_json = json.dumps(turns, indent=2)
+    except Exception as e:
+        # pyannote 3.4.0 (required for the gated 3.1 pipeline) needs an older
+        # torchaudio than this Space's torch stack ships (needed by Wan2.2 /
+        # LTX-Video / spaces itself) — diarization is best-effort here rather
+        # than something allowed to take down the whole call.
+        print(f"WARN: diarization unavailable: {e}", flush=True)
+        diarization_json = json.dumps(
+            {"error": "diarization unavailable on this Space (torch version conflict with pyannote)"}
+        )
 
     return transcript_text, diarization_json
 

requirements.txt

@@ -6,13 +6,11 @@ Pillow>=10.0
 scipy>=1.11
 imageio>=2.34
 imageio-ffmpeg>=0.5
-torch==2.2.2
-torchvision==0.17.2
-torchaudio==2.2.2
+torch>=2.2
+torchvision>=0.17
 diffusers>=0.38.0
 transformers>=4.40
 accelerate>=0.30
 peft>=0.10
 spaces>=0.30
 sentencepiece>=0.2
-pyannote.audio==3.4.0