Update services/model_service.py

services/model_service.py

@@ -34,6 +34,7 @@ class ModelService:
 
             # Load model with the updated configuration
             self.model = AutoModelForCausalLM.from_pretrained(
+                model_type == "llama" ,
                 settings.MODEL_NAME, 
                 torch_dtype=torch.float16 if settings.DEVICE == "cuda" else torch.float32,
                 device_map="auto" if settings.DEVICE == "cuda" else None