Implement proper adapter switching instead of multiple models

backend/models/character_manager.py

@@ -259,21 +259,21 @@ class CharacterManager:
                         with open(temp_config_file, 'w') as f:
                             json.dump(config_data, f, indent=2)
                         
-                        # Create a fresh model instance for this character to avoid adapter conflicts
-                        character_model = AutoModelForCausalLM.from_pretrained(
-                            self.base_model.name_or_path if hasattr(self.base_model, 'name_or_path') else settings.BASE_MODEL,
-                            torch_dtype=torch.float32,
-                            trust_remote_code=True,
-                            device_map="cpu"  # Keep on CPU to avoid memory issues
-                        )
-                        
-                        model_with_adapter = PeftModel.from_pretrained(
-                            character_model,
-                            temp_dir,
-                            adapter_name=character_id,
-                            is_trainable=False,
-                            torch_dtype=torch.float32,
-                        )
+                        # Use single model with adapter switching approach
+                        if not hasattr(self, 'peft_model'):
+                            # First adapter - create the PEFT model
+                            self.peft_model = PeftModel.from_pretrained(
+                                self.base_model,
+                                temp_dir,
+                                adapter_name=character_id,
+                                is_trainable=False,
+                                torch_dtype=torch.float32,
+                            )
+                            model_with_adapter = self.peft_model
+                        else:
+                            # Subsequent adapters - load as additional adapters
+                            self.peft_model.load_adapter(temp_dir, adapter_name=character_id)
+                            model_with_adapter = self.peft_model
                     
                     self.character_models[character_id] = model_with_adapter
                     logger.info(f"✅ Successfully loaded LoRA adapter for {character_id} with dedicated model instance")
@@ -333,6 +333,15 @@ class CharacterManager:
             
         # Get character-specific model and prompt
         model = self.character_models[character_id]
+        
+        # If using PEFT model with multiple adapters, switch to the correct one
+        if hasattr(self, 'peft_model') and hasattr(self.peft_model, 'set_adapter'):
+            try:
+                self.peft_model.set_adapter(character_id)
+                model = self.peft_model
+            except Exception as e:
+                logger.warning(f"Failed to switch adapter to {character_id}: {e}")
+                
         system_prompt = self.character_prompts.get(character_id, "")
         
         # Build conversation context