diff --git a/onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc b/onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc
index cc9068db28..1421ae1666 100644
--- a/onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc
+++ b/onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc
@@ -760,8 +760,8 @@ Status GreedySearchProcessLogits(
       cuda_stream);
 
 #ifdef DEBUG_GENERATION
-  dumper->Print("topk_scores", greedy_state->topk_scores_buffer.data(), batch_size);
-  dumper->Print("topk_indices", greedy_state->topk_tokens_buffer.data(), batch_size);
+  dumper->Print("topk_scores", greedy_state->topk_scores_buffer.data(), batch_size, 1);
+  dumper->Print("topk_indices", greedy_state->topk_tokens_buffer.data(), batch_size, 1);
 #endif
 
   CUDA_RETURN_IF_ERROR(cudaMemcpyAsync(greedy_state->next_tokens.data(),