diff --git a/src/coll/allreduce/cail_allreduce_rabenseifner.c b/src/coll/allreduce/cail_allreduce_rabenseifner.c
index 9a1dc32..f35016e 100644
--- a/src/coll/allreduce/cail_allreduce_rabenseifner.c
+++ b/src/coll/allreduce/cail_allreduce_rabenseifner.c
@@ -259,6 +259,10 @@ int cail_allreduce_rabenseifner(const void *sendbuf, void *recvbuf, int count,
             rc = PMPI_Send(recvbuf, count, datatype, rank - 1, 0, comm);
         } else {
             rc = PMPI_Recv(recvbuf, count, datatype, rank + 1, 0, comm, MPI_STATUS_IGNORE);
+            if (rc == MPI_SUCCESS) {
+                /* Flush the GPU-aware receive before the caller reads recvbuf. */
+                cail_gpu_flush_recv_buf(recvbuf, bufsize);
+            }
         }
     }
 
diff --git a/src/coll/allreduce/cail_allreduce_recursive_doubling.c b/src/coll/allreduce/cail_allreduce_recursive_doubling.c
index c4766ca..e028126 100644
--- a/src/coll/allreduce/cail_allreduce_recursive_doubling.c
+++ b/src/coll/allreduce/cail_allreduce_recursive_doubling.c
@@ -130,6 +130,8 @@ int cail_allreduce_recursive_doubling(const void *sendbuf, void *recvbuf,
         if ((rank % 2) == 0) {
             rc = PMPI_Recv(recvbuf, count, datatype, rank + 1, 0, comm, MPI_STATUS_IGNORE);
             if (rc != MPI_SUCCESS) goto cleanup;
+            /* Flush the GPU-aware receive before the caller reads recvbuf. */
+            cail_gpu_flush_recv_buf(recvbuf, bufsize);
         } else {
             rc = PMPI_Send(recvbuf, count, datatype, rank - 1, 0, comm);
             if (rc != MPI_SUCCESS) goto cleanup;