From 54de8024d3aa6fc4d85e0f1dff976d21e3eb5da9 Mon Sep 17 00:00:00 2001 From: Donato Capitella Date: Sun, 1 Feb 2026 11:14:25 +0000 Subject: [PATCH] Perf: Add NO_COMPRESS option to disable slow offload-compress --- projects/rccl/CMakeLists.txt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/projects/rccl/CMakeLists.txt b/projects/rccl/CMakeLists.txt index b46b27ec83..767b4a54da 100644 --- a/projects/rccl/CMakeLists.txt +++ b/projects/rccl/CMakeLists.txt @@ -1217,12 +1217,12 @@ if (HAVE_PARALLEL_JOBS) target_compile_options(rccl PRIVATE -parallel-jobs=12) endif() -if (ROCM_VERSION VERSION_GREATER_EQUAL "60200") +if (ROCM_VERSION VERSION_GREATER_EQUAL "60200" AND NOT NO_COMPRESS) target_compile_options(rccl PRIVATE --offload-compress) # Compress GPU code at compile time. target_link_libraries(rccl PRIVATE --offload-compress) # Compress GPU code at link time. message(STATUS "--offload-compress enabled - ROCm version >= 6.2.0") else() - message(STATUS "--offload-compress disabled - ROCm version < 6.2.0") + message(STATUS "--offload-compress disabled (ROCM < 6.2.0 or NO_COMPRESS=ON)") endif() target_compile_options(rccl PRIVATE -Werror=uninitialized)