Dynamically loading libnvidia-ml.so.1 instead of directly linking (#…

…313)

Currently, we install the driver into one of the CI images to allow for stub generation during compilation. This was needed because `device_info.cpp` linked directly against `CUDA::nvml`. This caused a link dependency on a driver library which is problematic when building with CPU-only docker images.

Instead, we dynamically load `libnvidia-ml.so.1` (appending the `.so.1` to avoid collisions with the stub file `libnvidia-ml.so`) and the necessary functions at runtime. If the library is not found, using a GPU will be disabled. This allows loading of the library for stub generation without needing a GPU.

Authors:
  - Michael Demoret (https://github.com/mdemoret-nv)

Approvers:
  - David Gardner (https://github.com/dagardner-nv)
  - Devin Robison (https://github.com/drobison00)

URL: #313

.github/workflows/ci_pipe.yml

@@ -132,7 +132,7 @@ jobs:
   test:
     name: Test
     needs: [build]
-    runs-on: [self-hosted, linux, amd64, gpu-v100-525-1]
+    runs-on: [self-hosted, linux, amd64, gpu-v100-latest-1]
     timeout-minutes: 60
     container:
       credentials:
@@ -171,7 +171,7 @@ jobs:
 
   codecov:
     name: Code Coverage
-    runs-on: [self-hosted, linux, amd64, gpu-v100-525-1]
+    runs-on: [self-hosted, linux, amd64, gpu-v100-latest-1]
     timeout-minutes: 60
     container:
       credentials:

.github/workflows/pull_request.yml

@@ -48,8 +48,8 @@ jobs:
     with:
       run_check: ${{ startsWith(github.ref_name, 'pull-request/') }}
       run_package_conda: ${{ !startsWith(github.ref_name, 'pull-request/') }}
-      container: nvcr.io/ea-nvidia-morpheus/morpheus:mrc-ci-driver-230410
-      test_container: nvcr.io/ea-nvidia-morpheus/morpheus:mrc-ci-test-230410
+      container: nvcr.io/ea-nvidia-morpheus/morpheus:mrc-ci-build-230412
+      test_container: nvcr.io/ea-nvidia-morpheus/morpheus:mrc-ci-test-230412
     secrets:
       CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
       CONDA_TOKEN: ${{ secrets.CONDA_TOKEN }}

Dockerfile

@@ -55,16 +55,11 @@ ENV CMAKE_CUDA_COMPILER_LAUNCHER=
 ENV CMAKE_CXX_COMPILER_LAUNCHER=
 ENV CMAKE_C_COMPILER_LAUNCHER=
 
-# ============ driver ==================
-FROM base as driver
+# ============ build ==================
+FROM base as build
 
-RUN --mount=type=cache,target=/var/cache/apt \
-    apt update && \
-    DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC \
-    apt install --no-install-recommends -y \
-    libnvidia-compute-525 \
-    && \
-    rm -rf /var/lib/apt/lists/*
+# Add any build only dependencies here. For now there is none but we need the
+# target to get the CI runner build scripts to work
 
 # ============ test ==================
 FROM base as test

cpp/mrc/CMakeLists.txt

@@ -166,7 +166,6 @@ target_link_libraries(libmrc
     mrc_protos
     mrc_architect_protos
     rmm::rmm
-    CUDA::nvml
     CUDA::cudart
     rxcpp::rxcpp
     glog::glog
@@ -180,8 +179,8 @@ target_link_libraries(libmrc
   PRIVATE
     hwloc::hwloc
     prometheus-cpp::core # private in MR !199
-    ucx::ucs
     ucx::ucp
+    ucx::ucs
 )
 
 target_include_directories(libmrc
@@ -206,10 +205,6 @@ target_compile_features(libmrc PUBLIC cxx_std_20)
 
 set_target_properties(libmrc PROPERTIES OUTPUT_NAME ${PROJECT_NAME})
 
-# Finally, set the install RPATH to include the stubs folder for CUDA::nvml. If thats made private, this can be removed
-set_target_properties(libmrc PROPERTIES INSTALL_RPATH
-    "${CMAKE_INSTALL_PREFIX}/lib:\$ORIGIN:${CMAKE_INSTALL_PREFIX}/lib/stubs")
-
 # ##################################################################################################
 # - install targets --------------------------------------------------------------------------------
 
@@ -277,4 +272,3 @@ rapids_export(BUILD ${PROJECT_NAME}
   DOCUMENTATION doc_string
   FINAL_CODE_BLOCK code_string
 )
-