Upd models loader #5

config.py

@@ -54,7 +54,8 @@ DESCRIPTION = """
 <p>📄 <strong>Document RAG:</strong> Answer based on uploaded medical documents</p>
 <p>🌐 <strong>Web Search:</strong> Fetch knowledge from reliable online medical resources</p>
 <p>🌍 <strong>Multi-language:</strong> Automatic translation for non-English queries</p>
-<p>Upload PDF or text files to get started!</p>
+<p>Tips: Customise configurations, system prompt to see where the magic happens!</p>
+<p>Note: Case GPU aborted errors, please select another model or try again later!</p>
 </center>
 """
 CSS = """

ui.py

@@ -356,7 +356,7 @@ def create_demo():
                             if config.global_whisper_model is not None:
                                 status_lines.append("✅ ASR (Whisper): loaded and ready")
                             else:
-                                status_lines.append("⚠️ ASR (Whisper): not loaded")
+                                status_lines.append("⏳ ASR (Whisper): will load on first use")
                         else:
                             status_lines.append("❌ ASR: library not available")
                         
@@ -394,9 +394,9 @@ def create_demo():
                     # ASR (Whisper) model status
                     if WHISPER_AVAILABLE:
                         if config.global_whisper_model is not None:
-                            status_lines.append("✅ ASR (Whisper large-v3-turbo): loaded and ready")
+                            status_lines.append("✅ ASR (Whisper): loaded and ready")
                         else:
-                            status_lines.append("⚠️ ASR (Whisper large-v3-turbo): not loaded")
+                            status_lines.append("⏳ ASR (Whisper): will load on first use")
                     else:
                         status_lines.append("❌ ASR: library not available")
                     
@@ -566,10 +566,44 @@ def create_demo():
                     outputs=[model_status, submit_button, message_input]
                 )
                 
-                # Load ALL models sequentially in a SINGLE GPU session to avoid ZeroGPU conflicts
-                # This prevents "GPU aborted" errors from multiple simultaneous GPU requests
+                # GPU-decorated function to load Whisper ASR model on-demand
+                @spaces.GPU(max_duration=120)
+                def load_whisper_model_on_demand():
+                    """Load Whisper ASR model when needed"""
+                    try:
+                        if WHISPER_AVAILABLE and config.global_whisper_model is None:
+                            logger.info("[ASR] Loading Whisper model on-demand...")
+                            initialize_whisper_model()
+                            if config.global_whisper_model is not None:
+                                logger.info("[ASR] ✅ Whisper model loaded successfully!")
+                                return "✅ ASR (Whisper): loaded"
+                            else:
+                                logger.warning("[ASR] ⚠️ Whisper model failed to load")
+                                return "⚠️ ASR (Whisper): failed to load"
+                        elif config.global_whisper_model is not None:
+                            return "✅ ASR (Whisper): already loaded"
+                        else:
+                            return "❌ ASR: library not available"
+                    except Exception as e:
+                        logger.error(f"[ASR] Error loading Whisper model: {e}")
+                        return f"❌ ASR: error - {str(e)[:100]}"
+                
+                # Load models on startup - medical model loads via stream_chat, Whisper loads on-demand
+                # Note: We skip startup loading to avoid GPU conflicts, models load when first needed
+                def update_startup_status():
+                    """Update status display on startup without loading models"""
+                    try:
+                        result = check_model_status(DEFAULT_MEDICAL_MODEL)
+                        if result and isinstance(result, tuple) and len(result) == 2:
+                            return result[0]
+                        else:
+                            return "⚠️ Checking model status..."
+                    except Exception as e:
+                        logger.error(f"Error in update_startup_status: {e}")
+                        return f"⚠️ Error: {str(e)[:100]}"
+                
                 demo.load(
-                    fn=load_all_models_on_startup,
+                    fn=update_startup_status,
                     inputs=None,
                     outputs=[model_status]
                 )

voice.py

@@ -81,11 +81,22 @@ def transcribe_audio_whisper(audio_path: str) -> str:
     try:
         logger.info(f"[ASR] Starting Whisper transcription for: {audio_path}")
         if config.global_whisper_model is None:
-            logger.info("[ASR] Whisper model not loaded, initializing...")
-            initialize_whisper_model()
+            logger.info("[ASR] Whisper model not loaded, initializing now (on-demand)...")
+            try:
+                initialize_whisper_model()
+                if config.global_whisper_model is None:
+                    logger.error("[ASR] Failed to initialize Whisper model - check logs for errors")
+                    return ""
+                else:
+                    logger.info("[ASR] ✅ Whisper model loaded successfully on-demand!")
+            except Exception as e:
+                logger.error(f"[ASR] Error initializing Whisper model: {e}")
+                import traceback
+                logger.debug(f"[ASR] Full traceback: {traceback.format_exc()}")
+                return ""
         
         if config.global_whisper_model is None:
-            logger.error("[ASR] Failed to initialize Whisper model")
+            logger.error("[ASR] Whisper model is still None after initialization attempt")
             return ""
         
         # Extract processor and model from stored dict