Try faster-whisper first for audio pre-load, fall back to GGUF

coderai

@@ -3613,17 +3613,13 @@ def main():
         if should_preload:
             print(f"Pre-loading audio model...")
             try:
+                # Try faster-whisper first (requires torch)
+                from faster_whisper import WhisperModel
+                import torch
+                
                 model_to_use = args.audio_model
                 model_path = None
                 
-                # Check if model is a GGUF file - use llama.cpp (Vulkan) instead of faster-whisper
-                is_gguf = model_to_use.endswith('.gguf') or '.gguf?' in model_to_use or 'gguf' in model_to_use.lower()
-                
-                if is_gguf:
-                    # Use llama.cpp for GGUF audio models (works with Vulkan)
-                    print(f"Using GGUF format with llama.cpp (Vulkan)...")
-                    from llama_cpp import Llama
-                    
                 # Check if model is a URL - handle caching
                 if model_to_use.startswith('http://') or model_to_use.startswith('https://'):
                     cached_path = get_cached_model_path(model_to_use)
@@ -3636,22 +3632,29 @@ def main():
                         model_path = download_model(model_to_use, cache_dir)
                         model_to_use = model_path
                 
-                    # Load with llama.cpp (Vulkan)
-                    audio_model = Llama(
-                        model_path=model_to_use,
-                        n_gpu_layers=-1,  # All layers to GPU
-                        n_ctx=2048,
-                        verbose=False
+                # Determine compute type
+                compute_type = "float16" if torch.cuda.is_available() else "int8"
+                
+                # Load the model
+                whisper_model = WhisperModel(
+                    model_to_use,
+                    device="cuda" if torch.cuda.is_available() else "cpu",
+                    compute_type=compute_type
                 )
                 
                 # Store in multi_model_manager
                 model_key = f"audio:{args.audio_model}"
-                    multi_model_manager.add_model(model_key, audio_model)
-                    print(f"Audio model loaded successfully (GGUF/Vulkan)")
-                else:
-                    # Use faster-whisper for non-GGUF models
-                    from faster_whisper import WhisperModel
-                    import torch
+                multi_model_manager.add_model(model_key, whisper_model)
+                print(f"Audio model loaded successfully (faster-whisper)")
+                
+            except ImportError:
+                # faster-whisper not available, try GGUF with llama.cpp
+                print("faster-whisper not available, trying GGUF with llama.cpp...")
+                try:
+                    from llama_cpp import Llama
+                    
+                    model_to_use = args.audio_model
+                    model_path = None
                     
                     # Check if model is a URL - handle caching
                     if model_to_use.startswith('http://') or model_to_use.startswith('https://'):
@@ -3665,26 +3668,27 @@ def main():
                             model_path = download_model(model_to_use, cache_dir)
                             model_to_use = model_path
                     
-                    # Determine compute type
-                    compute_type = "float16" if torch.cuda.is_available() else "int8"
-                    
-                    # Load the model
-                    whisper_model = WhisperModel(
-                        model_to_use,
-                        device="cuda" if torch.cuda.is_available() else "cpu",
-                        compute_type=compute_type
+                    # Load with llama.cpp (Vulkan)
+                    audio_model = Llama(
+                        model_path=model_to_use,
+                        n_gpu_layers=-1,  # All layers to GPU
+                        n_ctx=2048,
+                        verbose=False
                     )
                     
                     # Store in multi_model_manager
                     model_key = f"audio:{args.audio_model}"
-                    multi_model_manager.add_model(model_key, whisper_model)
-                    print(f"Audio model loaded successfully")
+                    multi_model_manager.add_model(model_key, audio_model)
+                    print(f"Audio model loaded successfully (GGUF/Vulkan)")
                     
                 except Exception as e:
                     print(f"Warning: Could not pre-load audio model: {e}")
                     import traceback
                     traceback.print_exc()
             
+            except Exception as e:
+                print(f"Warning: Could not pre-load audio model: {e}")
+    
     # Set up TTS model if specified
     if args.tts_model:
         print(f"\nText-to-speech model: {args.tts_model}")