diff --git a/cpp/src/sampling/neighbor_sampling_impl.hpp b/cpp/src/sampling/neighbor_sampling_impl.hpp
index feaa8824d73..bde467b9407 100644
--- a/cpp/src/sampling/neighbor_sampling_impl.hpp
+++ b/cpp/src/sampling/neighbor_sampling_impl.hpp
@@ -200,8 +200,14 @@ neighbor_sample_impl(
   std::vector<size_t> level_sizes{};
   int32_t hop{0};
   int32_t edge_type_id_max{1}; // A value of 1 translate to homogeneous neighbor sample
+  int32_t num_edge_type_per_hop{0};
   
   auto cur_graph_view = modified_graph_view ? *modified_graph_view : graph_view;
+
+  if (heterogeneous_fan_out) {
+    num_edge_type_per_hop = std::get<0>(*heterogeneous_fan_out).back() - 1;
+  }
+
   while(true) {
     int32_t k_level{0};
     if (fan_out) {
@@ -210,8 +216,11 @@ neighbor_sample_impl(
         break;
       }
     } else if (heterogeneous_fan_out) {
-      // initially edge type
-      edge_type_id_max = std::get<0>(*heterogeneous_fan_out).back() - 1;
+        if (num_edge_type_per_hop == 0) {
+          break;
+        }
+        edge_type_id_max = std::get<0>(*heterogeneous_fan_out).back() - 1;
+      
     }
   
     for (int i = 0; i < edge_type_id_max; i++) {
@@ -223,7 +232,10 @@ neighbor_sample_impl(
         auto k_level_size = (std::get<1>(*heterogeneous_fan_out)[i + 1] - std::get<1>(*heterogeneous_fan_out)[i]);
         if (k_level_size > hop) {
           k_level = i + hop;
-        } // otherwise, k_level = 0
+        } else { // otherwise, k_level = 0
+          --num_edge_type_per_hop ;
+
+        }
       }
       rmm::device_uvector<vertex_t> srcs(0, handle.get_stream());
       rmm::device_uvector<vertex_t> dsts(0, handle.get_stream());