Check nvidia_peermem during runtime

.azure-pipelines/integration-test.yml

@@ -31,7 +31,7 @@ jobs:
       targetType: 'inline'
       script: |
         mkdir build && cd build
-        cmake -DCMAKE_BUILD_TYPE=Release -DBYPASS_PEERMEM_CHECK=ON -DBYPASS_GPU_CHECK=ON -DUSE_CUDA=ON ..
+        cmake -DCMAKE_BUILD_TYPE=Release -DBYPASS_GPU_CHECK=ON -DUSE_CUDA=ON ..
         make -j
       workingDirectory: '$(System.DefaultWorkingDirectory)'
 

.azure-pipelines/multi-nodes-test.yml

@@ -26,7 +26,7 @@ jobs:
       targetType: 'inline'
       script: |
         mkdir build && cd build
-        cmake -DCMAKE_BUILD_TYPE=Release -DBYPASS_PEERMEM_CHECK=ON -DBYPASS_GPU_CHECK=ON -DUSE_CUDA=ON ..
+        cmake -DCMAKE_BUILD_TYPE=Release -DBYPASS_GPU_CHECK=ON -DUSE_CUDA=ON ..
         make -j
         make pylib-copy
       workingDirectory: '$(System.DefaultWorkingDirectory)'

.github/workflows/codeql-analysis.yml

@@ -45,7 +45,7 @@ jobs:
 
     - name: Build
       run: |
-        cmake -DBYPASS_PEERMEM_CHECK=ON -DBYPASS_GPU_CHECK=ON -DUSE_CUDA=ON .
+        cmake -DBYPASS_GPU_CHECK=ON -DUSE_CUDA=ON .
         make -j
 
     - name: Perform CodeQL Analysis

CMakeLists.txt

@@ -21,7 +21,6 @@ option(BUILD_PYTHON_BINDINGS "Build Python bindings" ON)
 option(USE_CUDA "Use NVIDIA/CUDA." OFF)
 option(USE_ROCM "Use AMD/ROCm." OFF)
 option(BYPASS_GPU_CHECK "Bypass GPU check." OFF)
-option(BYPASS_PEERMEM_CHECK "Bypass checking nvidia_peermem" OFF)
 
 if(BYPASS_GPU_CHECK)
     if(USE_CUDA)
@@ -81,16 +80,6 @@ if(USE_CUDA)
 
     set(GPU_LIBRARIES CUDA::cudart CUDA::cuda_driver)
     set(GPU_INCLUDE_DIRS ${CUDAToolkit_INCLUDE_DIRS})
-
-    # Find if nvidia_peermem is installed and loaded
-    if(NOT BYPASS_PEERMEM_CHECK)
-        execute_process(COMMAND sh -c "lsmod | grep nvidia_peermem"
-                        RESULT_VARIABLE lsmod_result
-                        OUTPUT_VARIABLE lsmod_output)
-        if(NOT lsmod_result EQUAL 0)
-            message(FATAL_ERROR "nvidia_peermem is not installed or not loaded.")
-        endif()
-    endif()
 else()
     set(CMAKE_HIP_STANDARD 17)
     set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Wall -Wextra")

src/ib.cc

@@ -8,6 +8,7 @@
 #include <unistd.h>
 
 #include <cstring>
+#include <fstream>
 #include <mscclpp/core.hpp>
 #include <mscclpp/fifo.hpp>
 #include <sstream>
@@ -16,6 +17,20 @@
 #include "api.h"
 #include "debug.h"
 
+#if !defined(__HIP_PLATFORM_AMD__) || (__HIP_PLATFORM_AMD__ == 0)
+
+// Check if nvidia_peermem kernel module is loaded
+static bool checkNvPeerMemLoaded() {
+  std::ifstream file("/proc/modules");
+  std::string line;
+  while (std::getline(file, line)) {
+    if (line.find("nvidia_peermem") != std::string::npos) return true;
+  }
+  return false;
+}
+
+#endif  // !defined(__HIP_PLATFORM_AMD__) || (__HIP_PLATFORM_AMD__ == 0)
+
 namespace mscclpp {
 
 IbMr::IbMr(ibv_pd* pd, void* buff, std::size_t size) : buff(buff) {
@@ -280,6 +295,11 @@ const ibv_wc* IbQp::getWc(int idx) const { return &this->wcs[idx]; }
 int IbQp::getNumCqItems() const { return this->numSignaledPostedItems; }
 
 IbCtx::IbCtx(const std::string& devName) : devName(devName) {
+#if !defined(__HIP_PLATFORM_AMD__) || (__HIP_PLATFORM_AMD__ == 0)
+  if (!checkNvPeerMemLoaded()) {
+    throw mscclpp::Error("nvidia_peermem kernel module is not loaded", ErrorCode::InternalError);
+  }
+#endif  // !defined(__HIP_PLATFORM_AMD__) || (__HIP_PLATFORM_AMD__ == 0)
   int num;
   struct ibv_device** devices = ibv_get_device_list(&num);
   for (int i = 0; i < num; ++i) {