diff --git a/include/nccl_ofi_rdma.h b/include/nccl_ofi_rdma.h
index f0206d945..7cacc3dff 100644
--- a/include/nccl_ofi_rdma.h
+++ b/include/nccl_ofi_rdma.h
@@ -5,9 +5,7 @@
 #ifndef NCCL_OFI_RDMA_H_
 #define NCCL_OFI_RDMA_H_
 
-#ifdef __cplusplus
-extern "C" {
-#endif
+#include "config.h"
 
 #include <uthash/uthash.h>
 #include <rdma/fabric.h>
@@ -21,7 +19,9 @@ extern "C" {
 #include "nccl_ofi_msgbuff.h"
 #include "nccl_ofi_scheduler.h"
 #include "nccl_ofi_topo.h"
-#include "nccl_ofi_tracepoint.h"
+#if HAVE_NVTX_TRACING
+#include <nvtx3/nvToolsExt.h>
+#endif
 
 /* Maximum number of rails supported. This defines the size of
  * messages exchanged during connection establishment (linear
@@ -840,6 +840,9 @@ typedef struct nccl_net_ofi_rdma_device {
 #endif
 } nccl_net_ofi_rdma_device_t;
 
+#ifdef __cplusplus
+extern "C" {
+#endif
 
 struct nccl_net_ofi_rdma_plugin {
 	nccl_net_ofi_plugin_t base;