diff --git a/modelscope/pipelines/nlp/llm_pipeline.py b/modelscope/pipelines/nlp/llm_pipeline.py
index f257a8f0..c69d0c5b 100644
--- a/modelscope/pipelines/nlp/llm_pipeline.py
+++ b/modelscope/pipelines/nlp/llm_pipeline.py
@@ -108,8 +108,6 @@ class LLMPipeline(Pipeline):
         is_messages = isinstance(inputs, dict) and 'messages' in inputs
         tokens = self.preprocess(inputs, is_messages, **preprocess_params)
 
-        response = dict()
-
         if output_logits:
             with torch.no_grad():
                 if hasattr(self.model, 'model'):
@@ -117,6 +115,7 @@ class LLMPipeline(Pipeline):
                 else:
                     outputs = self.model(
                         tokens['inputs'])  # [batch, seq, vocab]
+                return {'logits': outputs[0], 'tokens': tokens['inputs']}
         else:
             if hasattr(self.model, 'generate'):
                 outputs = self.model.generate(**tokens, **forward_params)
@@ -130,16 +129,7 @@ class LLMPipeline(Pipeline):
             response = self.postprocess(outputs, is_messages,
                                         **postprocess_params)
 
-        if output_logits:
-            logits_d: dict = {
-                'logits': outputs[0].cpu(),
-                'tokens': tokens['inputs'].cpu(),
-                'inputs_len': len(tokens['inputs'][0]),
-                'decode_func': self.tokenizer.decode
-            }
-            response.update(logits_d)
-
-        return response
+            return response
 
     def preprocess(self, inputs: Union[str, Dict], is_messages: bool,
                    **kwargs):