diff --git a/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.cu b/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.cu
index 45ed087aac..5e1f14c547 100644
--- a/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.cu
+++ b/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.cu
@@ -434,6 +434,10 @@ void LaunchBeamSearchScorer_Process(BeamScorerState& state_cpu,
                                     gsl::span<const int32_t> next_tokens,
                                     gsl::span<const int32_t> next_indices,
                                     cudaStream_t stream) {
+  for (size_t i = 0; i < beam_hyps.length(); i++) {
+    dump_beam_hypotheses(&beam_hyps[i]);
+  }
+
   BeamSearchScorer_Process<<<1, state_cpu.batch_size_, 0, stream>>>(state_cpu,
                                                                     state,
                                                                     sequences.data(),
diff --git a/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.h b/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.h
index 0a68a822c7..3bf8ef34b3 100644
--- a/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.h
+++ b/onnxruntime/contrib_ops/cuda/transformers/generation_cuda_impl.h
@@ -52,6 +52,37 @@ struct HypothesisScore {
   float score;
 };
 
+// Function to dump the data in the struct
+__device__ void dump_hypothesis_score(const struct HypothesisScore* hs) {
+    printf("HypothesisScore Dump:\n");
+    printf("  hypothesis_length: %d\n", hs->hypothesis_length);
+    printf("  score: %f\n", hs->score);
+
+    printf("  hypothesis: ");
+    if (hs->hypothesis_length > 0 && hs->hypothesis != NULL) {
+        for (int i = 0; i < hs->hypothesis_length; ++i) {
+            printf("%d ", hs->hypothesis[i]);
+        }
+    } else {
+        printf("(empty)");
+    }
+    printf("\n");
+}
+
+__device__ void dump_beam_hypotheses(const struct BeamHypotheses* bh) {
+    printf("BeamHypotheses Dump:\n");
+    printf("  beams_count_: %d\n", bh->beams_count_);
+    printf("  beams_used_: %d\n", bh->beams_used_);
+    printf("  length_penalty_: %f\n", bh->length_penalty_);
+    printf("  done_: %s\n", bh->done_ ? "true" : "false");
+
+    printf("  beams_:\n");
+    for (int i = 0; i < bh->beams_used_; ++i) {
+        printf("    Beam %d:\n", i + 1);
+        dump_hypothesis_score(&bh->beams_[i]);
+    }
+}
+
 struct BeamHypotheses {
   HypothesisScore* beams_;  // Beam width sized array of hypotheses, sorted by highest scoring
   int beams_count_;