diff --git a/src/llama-calrt.cpp b/src/llama-calrt.cpp
index 5bbdd52e1f7adc96ac72945ed52eabd9e075d990..e038a4c25d5589550680bf5bb556145f1524ae17 100644
--- a/src/llama-calrt.cpp
+++ b/src/llama-calrt.cpp
@@ -564,6 +564,10 @@ void calrt_context::calrt_infer(calrt_seqs_info & seqs) {
         // makeup_ubatch(ubatch);
         copy_data_to_ibuf(*pIbuf, ubatch, model_name);
 
+        //only copy promt tokens
+        //pObuf->SliceTensorByName(pObuf->GetTensorByName("outputs[0]")->GetDataPtr(),"outputs[0]", 0, cur_seq_len * sizeof(ggml_bf16_t));
+        pObuf->GetTensorByName("outputs[0]")->SliceTensor(0, cur_seq_len * sizeof(ggml_bf16_t));
+
         calrt::infer(vdev, model, pIbuf.get(), pObuf.get());
         pObuf->Wait();