Update app.py

app.py

@@ -1,62 +1,52 @@
 import os
 import numpy as np
 import torch
+import typing
 
-# --- CRITICAL ENVIRONMENT FIXES ---
-# 1. Fix for Hugging Face millicore OMP_NUM_THREADS error
-if os.environ.get("OMP_NUM_THREADS", "").endswith("m"):
-    os.environ["OMP_NUM_THREADS"] = "1"
-
-# 2. GLOBAL PYTORCH 2.6+ SECURITY BYPASS (MONKEYPATCH)
-# This forces torch.load to behave like older versions (trusting the file).
-# Essential for loading 3rd party checkpoints (WhisperX/Pyannote) that aren't updated yet.
+# --- CRITICAL: AGGRESSIVE PYTORCH SECURITY OVERRIDE ---
+# The error "Weights only load failed" means weights_only=True is active.
+# We must intercept torch.load and FORCE it to False for pyannote/whisperx to work.
 _original_torch_load = torch.load
+
 def patched_torch_load(*args, **kwargs):
-    if 'weights_only' not in kwargs:
-        kwargs['weights_only'] = False
+    # Debug print to verify interception (Check Logs tab in HF)
+    print(f"DEBUG: Intercepted torch.load call. Target: {args[0] if args else 'Unknown'}")
+    
+    # FORCE Disable security check
+    kwargs['weights_only'] = False 
+    
     return _original_torch_load(*args, **kwargs)
-torch.load = patched_torch_load
 
-# 3. Force Torchaudio Backend
-import torchaudio
-try:
-    if "ffmpeg" in torchaudio.list_audio_backends():
-        torchaudio.set_audio_backend("ffmpeg")
-except Exception:
-    pass
+# Apply the patch
+torch.load = patched_torch_load
+print("DEBUG: torch.load has been monkeypatched to allow all globals.")
 
-# --- SAFE GLOBALS WHITELIST (Backup Strategy) ---
-# Even with the monkeypatch, we add these to be double-safe.
-safe_globals = []
+# --- Secondary Backup: Safe Globals Whitelist ---
 try:
-    # PyTorch internals
-    safe_globals.append(torch.nn.modules.container.ModuleList)
-    
-    # NumPy internals
-    safe_globals.append(np.dtype)
-    if hasattr(np, '_core') and hasattr(np._core, 'multiarray'):
-        safe_globals.append(np._core.multiarray.scalar)
-    elif hasattr(np, 'core') and hasattr(np.core, 'multiarray'):
-        safe_globals.append(np.core.multiarray.scalar)
-
-    # OmegaConf (Used by Pyannote config)
-    import omegaconf
     from omegaconf.listconfig import ListConfig
     from omegaconf.dictconfig import DictConfig
-    from omegaconf.base import ContainerMetadata, Metadata, Node
-    safe_globals.extend([ListConfig, DictConfig, ContainerMetadata, Metadata, Node])
-    
-    # Pyannote internals (if available at this stage)
-    # We wrap these in try/except in case pyannote isn't fully loaded yet
     try:
-        from pyannote.audio.core.task import Specifications, Problem, Resolution
-        from pyannote.audio.core.model import Model
-        from pyannote.audio.pipelines.speaker_diarization import SpeakerDiarization
-        safe_globals.extend([Specifications, Problem, Resolution, Model, SpeakerDiarization])
+        from omegaconf.base import ContainerMetadata, Metadata
     except ImportError:
-        pass
+        ContainerMetadata = None
+        Metadata = None
 
-    torch.serialization.add_safe_globals(safe_globals)
+    safe_list = [
+        typing.Any, # Specific fix for your error
+        ListConfig, 
+        DictConfig,
+        torch.nn.modules.container.ModuleList,
+        np.dtype,
+    ]
+    if ContainerMetadata: safe_list.append(ContainerMetadata)
+    if Metadata: safe_list.append(Metadata)
+    
+    if hasattr(np, '_core') and hasattr(np._core, 'multiarray'):
+        safe_list.append(np._core.multiarray.scalar)
+    elif hasattr(np, 'core') and hasattr(np.core, 'multiarray'):
+        safe_list.append(np.core.multiarray.scalar)
+
+    torch.serialization.add_safe_globals(safe_list)
 except Exception as e:
     print(f"Safe Globals Warning: {e}")
 
@@ -83,6 +73,10 @@ ENV_GEMINI_KEY = os.environ.get("GEMINI_API_KEY", "")
 ACTIVE_HF_TOKEN = ENV_HF_TOKEN if ENV_HF_TOKEN else HARDCODED_HF_TOKEN
 ACTIVE_GEMINI_KEY = ENV_GEMINI_KEY if ENV_GEMINI_KEY else HARDCODED_GEMINI_KEY
 
+# Fix OMP Threads
+if os.environ.get("OMP_NUM_THREADS", "").endswith("m"):
+    os.environ["OMP_NUM_THREADS"] = "1"
+
 def format_timecode(seconds, fps=25):
     """Converts seconds to HH:MM:SS:FF."""
     td = timedelta(seconds=seconds)
@@ -156,20 +150,6 @@ with st.sidebar:
     
     st.header("Model Settings")
     model_size = st.selectbox("Whisper Model", ["large-v2", "medium", "base"], index=0)
-    
-    # --- New Language Option ---
-    language_map = {
-        "Auto-Detect": None,
-        "English": "en",
-        "Spanish": "es",
-        "French": "fr",
-        "German": "de",
-        "Italian": "it",
-        "Portuguese": "pt"
-    }
-    selected_lang_label = st.selectbox("Audio Language (Speeds up processing)", list(language_map.keys()), index=1)
-    target_language = language_map[selected_lang_label]
-    
     num_speakers = st.number_input("Speakers (0=Auto)", min_value=0, value=0)
     
     st.divider()
@@ -200,9 +180,9 @@ if uploaded_file:
                     try:
                         device = "cuda" if torch.cuda.is_available() else "cpu"
                         if device == "cpu":
-                            st.warning("⚠️ No GPU detected. WhisperX will be very slow.")
-                            
-                        st.write(f"🚀 **Loading WhisperX on {device}...**")
+                            st.warning("⚠️ No GPU detected. WhisperX will be slow.")
+                        else:
+                            st.write(f"🚀 **Loading WhisperX on {device}...**")
                         
                         # 1. Transcribe
                         compute_type = "float16" if device == "cuda" else "int8"
@@ -210,11 +190,8 @@ if uploaded_file:
                         
                         st.write("📝 **Transcribing...**")
                         audio = whisperx.load_audio("temp_audio.wav")
+                        result = model.transcribe(audio, batch_size=16)
                         
-                        # Pass the language to speed up processing
-                        result = model.transcribe(audio, batch_size=16, language=target_language)
-                        
-                        # Memory cleanup
                         del model
                         gc.collect()
                         torch.cuda.empty_cache()
@@ -230,6 +207,7 @@ if uploaded_file:
                         
                         # 3. Diarize
                         st.write("🗣️ **Diarizing Speakers...**")
+                        # The monkeypatch above should protect this call
                         diarize_model = whisperx.DiarizationPipeline(use_auth_token=ACTIVE_HF_TOKEN, device=device)
                         
                         diarize_kwargs = {}