diff --git a/paddle/fluid/operators/class_center_sample_op.cu b/paddle/fluid/operators/class_center_sample_op.cu
index 29286be0dd6b2082586299765310c7428365a87a..f1ccbc913d9b18c580cd9066f5ffa28d187e7507 100644
--- a/paddle/fluid/operators/class_center_sample_op.cu
+++ b/paddle/fluid/operators/class_center_sample_op.cu
@@ -390,18 +390,26 @@ class ClassCenterSampleCUDAKernel : public framework::OpKernel<T> {
         ctx.cuda_device_context().stream())));
 
     // step 5: random sample negative class center
+    uint64_t seed_data;
+    uint64_t increment;
     int vec_size = VectorizedSize<T>(cub_sort_keys_ptr);
-    int increment = ((num_classes - 1) /
-                         (NumBlocks(num_classes) * kNumCUDAThreads * vec_size) +
-                     1) *
-                    vec_size;
-    if (!fix_seed) {
+    auto offset = ((num_classes - 1) /
+                       (NumBlocks(num_classes) * kNumCUDAThreads * vec_size) +
+                   1) *
+                  vec_size;
+    auto gen_cuda = framework::GetDefaultCUDAGenerator(rank);
+    if (gen_cuda->GetIsInitPy() && (!fix_seed)) {
+      auto seed_offset = gen_cuda->IncrementOffset(offset);
+      seed_data = seed_offset.first;
+      increment = seed_offset.second;
+    } else {
       std::random_device rnd;
-      seed = rnd();
+      seed_data = fix_seed ? seed + rank : rnd();
+      increment = offset;
     }
     RandomSampleClassCenter<T><<<NumBlocks(num_classes), kNumCUDAThreads, 0,
                                  ctx.cuda_device_context().stream()>>>(
-        num_classes, seed + rank, increment, num_classes, cub_sort_keys_ptr);
+        num_classes, seed_data, increment, num_classes, cub_sort_keys_ptr);
 
     // step 6: mark positive class center as negative value
     // fill the sort values to index 0, 1, ..., batch_size-1