diff --git a/paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu b/paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu
index 91c583fff8549cd35ebfbfde832002563031f4ca..09233ab99574b865a66279126070f11fb8868d97 100644
--- a/paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu
+++ b/paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu
@@ -1403,7 +1403,7 @@ class DistributedFusedLambOpKernel<platform::CUDADeviceContext, T>
 
     // Step 6: allreduce + global norm gradient clip
     int64_t global_rank = 0, local_rank = 0;
-    ncclComm_t global_comm = nullptr, local_comm = 0;
+    ncclComm_t global_comm = nullptr, local_comm = nullptr;
     if (nranks > 1) {
       auto *nccl_comm_handle =
           platform::NCCLCommContext::Instance().Get(ring_ids[0], place);