Update handler.py

handler.py

@@ -19,7 +19,7 @@ class EndpointHandler:
             decoder_file_name="decoder_model_quantized.onnx",
             decoder_file_with_past_name="decoder_with_past_model_quantized.onnx",
         )
-        self.pipe = TranslationPipeline(self.model, self.tokenizer)
+        self.pipe = TranslationPipeline(self.model, self.tokenizer, do_sample=False, num_beams=1)
         self.translator = Translator(self.pipe, max_length=512, batch_size=1)
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]: