From 595cda2ab935f44749a7bbc8533a643238df7518 Mon Sep 17 00:00:00 2001 From: Abhishek Kulkarni <11399+adk9@users.noreply.github.com> Date: Sat, 2 Nov 2024 23:20:07 -0700 Subject: [PATCH] GDR enablement logic fix for kernel 6.4.0+ (#1378) [ROCm/rccl commit: 61785568535667bca70f871ce797824c52d16fac] --- projects/rccl/src/transport/net_ib.cc | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/projects/rccl/src/transport/net_ib.cc b/projects/rccl/src/transport/net_ib.cc index 6886e27fe5..6d77784422 100644 --- a/projects/rccl/src/transport/net_ib.cc +++ b/projects/rccl/src/transport/net_ib.cc @@ -607,7 +607,9 @@ ncclResult_t ncclIbGdrSupport() { NCCLCHECK(ncclTopoGetStrFromSys("/proc/sys/kernel", "numa_balancing", strValue)); if (strcmp(strValue, "1") == 0 && roMode == 0) moduleLoaded = 0; - } else if (moduleLoaded == 0) { + } + + if (moduleLoaded == 0) { // Check for `ib_register_peer_memory_client` symbol in `/proc/kallsyms` // if your system uses native OS ib_peer module char buf[256];