Update handler.py

handler.py

@@ -77,7 +77,7 @@ class EndpointHandler:
                     )
 
             generation_config = deepcopy(self.generation_config)
-            generation_config.update(**input.get("generation_parameters", {}))
+            generation_config.update(**input.get("generation_parameters", {"max_new_tokens": 128}))
 
             messages = [
                 {
@@ -100,6 +100,6 @@ class EndpointHandler:
                 generated_ids,
                 skip_special_tokens=True,
             )
-            predictions.append(generated_texts[0])
+            predictions.append(generated_texts[0][len(prompt):])
 
         return {"predictions": predictions}