Fix adapter interference: create dedicated model instances per character

backend/models/character_manager.py

@@ -259,8 +259,16 @@ 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(
-                            self.base_model,
+                            character_model,
                             temp_dir,
                             adapter_name=character_id,
                             is_trainable=False,
@@ -268,7 +276,7 @@ class CharacterManager:
                         )
                     
                     self.character_models[character_id] = model_with_adapter
-                    logger.info(f"✅ Successfully loaded LoRA adapter for {character_id} with cleaned config")
+                    logger.info(f"✅ Successfully loaded LoRA adapter for {character_id} with dedicated model instance")
                     
                     # Cleanup temp files
                     shutil.rmtree(temp_dir, ignore_errors=True)