[Release/2.3] Changes to support UB 24.04 build (#1633) (#1952)

Changes are based on branch by @iupaikov-amd -
https://github.com/ROCm/pytorch/tree/ypajkov/ubuntu_noble_build
Validation:
http://rocm-ci.amd.com/view/Release-6.4/job/framework-pytorch-2.3-ub24-py3.12-ci_rel-6.4/3/

Fixes https://ontrack-internal.amd.com/browse/SWDEV-520708

Co-authored-by: Pruthvi Madugundu <[email protected]>

Author: ethanwee1

.ci/docker/common/install_base.sh

@@ -15,6 +15,9 @@ install_ubuntu() {
   elif [[ "$UBUNTU_VERSION" == "22.04"* ]]; then
     cmake3="cmake=3.22*"
     maybe_libiomp_dev=""
+  elif [[ "$UBUNTU_VERSION" == "24.04"* ]]; then
+    cmake3="cmake=3.28*"
+    maybe_libiomp_dev=""
   else
     cmake3="cmake=3.5*"
     maybe_libiomp_dev="libiomp-dev"

.ci/docker/common/install_conda.sh

@@ -118,6 +118,10 @@ fi
     conda_install_through_forge libstdcxx-ng=12
   fi
 
+  if [ "$ANACONDA_PYTHON_VERSION" = "3.12" ] || [ "$UBUNTU_VERSION" == "24.04"* ] ; then
+    conda_install_through_forge libstdcxx-ng=14
+  fi
+
   # Install some other packages, including those needed for Python test reporting
   pip_install -r /opt/conda/requirements-ci.txt
 

.ci/docker/common/install_rocm.sh

@@ -19,7 +19,7 @@ install_ubuntu() {
       # gpg-agent is not available by default on 20.04
       apt-get install -y --no-install-recommends gpg-agent
     fi
-    if [[ $UBUNTU_VERSION == 22.04 ]]; then
+    if [[ $UBUNTU_VERSION == 22.04 ]] || [[ $UBUNTU_VERSION == 24.04 ]]; then
         apt-get install -y --no-install-recommends gpg-agent
         echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' \
             | sudo tee /etc/apt/preferences.d/rocm-pin-600

.ci/docker/common/install_user.sh

@@ -2,6 +2,13 @@
 
 set -ex
 
+# Since version 24 the system ships with user 'ubuntu' that has id 1000
+# We need a work-around to enable id 1000 usage for this script
+if [[ $UBUNTU_VERSION == 24.04 ]]; then
+    # touch is used to disable harmless error message
+    touch /var/mail/ubuntu && chown ubuntu /var/mail/ubuntu && userdel -r ubuntu
+fi
+
 # Mirror jenkins user in container
 # jenkins user as ec2-user should have the same user-id
 echo "jenkins:x:1000:1000::/var/lib/jenkins:" >> /etc/passwd

.ci/docker/requirements-ci.txt

@@ -22,6 +22,7 @@ sympy==1.12.1
 #test that import:
 
 coremltools==5.0b5 ; python_version < "3.12"
+coremltools==7.2 ; python_version == "3.12"
 #Description: Apple framework for ML integration
 #Pinned versions: 5.0b5
 #test that import:
@@ -64,6 +65,7 @@ lark==0.12.0
 #test that import:
 
 librosa>=0.6.2 ; python_version < "3.11"
+librosa==0.10.2 ; python_version == "3.12"
 #Description: A python package for music and audio analysis
 #Pinned versions: >=0.6.2
 #test that import: test_spectral_ops.py
@@ -112,6 +114,7 @@ networkx==2.8.8
 numba==0.49.0 ; python_version < "3.9"
 numba==0.54.1 ; python_version == "3.9"
 numba==0.55.2 ; python_version == "3.10"
+numba==0.60.0 ; python_version == "3.12"
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
 #test that import: test_numba_integration.py

c10/hip/CMakeLists.txt

@@ -44,7 +44,7 @@ endif()
 # ---[ Dependency of c10_hip
 target_link_libraries(c10_hip PUBLIC c10)
 
-target_link_libraries(c10_hip PUBLIC ${PYTORCH_HIP_LIBRARIES})
+  target_link_libraries(c10_hip PUBLIC ${PYTORCH_HIP_LIBRARIES} ${ROCM_HSART_LIB})
 
 target_include_directories(
     c10_hip PUBLIC

caffe2/CMakeLists.txt

@@ -723,7 +723,7 @@ if(USE_ROCM)
   # caffe2_nvrtc's stubs to driver APIs are useful for HIP.
   # See NOTE [ ATen NVRTC Stub and HIP ]
   add_library(caffe2_nvrtc SHARED ${ATen_NVRTC_STUB_SRCS})
-  target_link_libraries(caffe2_nvrtc ${PYTORCH_HIP_LIBRARIES} ${ROCM_HIPRTC_LIB})
+  target_link_libraries(caffe2_nvrtc ${PYTORCH_HIP_LIBRARIES} ${ROCM_HIPRTC_LIB} ${ROCM_HSART_LIB})
   target_include_directories(caffe2_nvrtc PRIVATE ${CMAKE_BINARY_DIR})
   target_compile_definitions(caffe2_nvrtc PRIVATE USE_ROCM __HIP_PLATFORM_AMD__)
   install(TARGETS caffe2_nvrtc DESTINATION "${TORCH_INSTALL_LIB_DIR}")
@@ -1451,6 +1451,7 @@ target_link_libraries(torch_cpu PUBLIC c10)
 target_link_libraries(torch_cpu PUBLIC ${Caffe2_PUBLIC_DEPENDENCY_LIBS})
 target_link_libraries(torch_cpu PRIVATE ${Caffe2_DEPENDENCY_LIBS})
 target_link_libraries(torch_cpu PRIVATE ${Caffe2_DEPENDENCY_WHOLE_LINK_LIBS})
+target_link_libraries(torch_cpu PUBLIC ${ROCM_HSART_LIB})
 if(USE_MPI)
   target_link_libraries(torch_cpu PRIVATE MPI::MPI_CXX)
 endif()

cmake/Dependencies.cmake

@@ -1332,11 +1332,8 @@ if(USE_ROCM)
     hip_include_directories(${Caffe2_HIP_INCLUDE})
 
     set(Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS
-      ${PYTORCH_HIP_LIBRARIES} ${PYTORCH_MIOPEN_LIBRARIES} ${hipcub_LIBRARIES} ${ROCM_HIPRTC_LIB} ${ROCM_ROCTX_LIB})
-    if(ROCM_VERSION_DEV VERSION_GREATER_EQUAL "5.7.0")
-      list(APPEND Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS ${hipblaslt_LIBRARIES})
-    endif()
-
+      ${PYTORCH_HIP_LIBRARIES} ${PYTORCH_MIOPEN_LIBRARIES} ${hipcub_LIBRARIES} ${ROCM_HIPRTC_LIB} ${ROCM_ROCTX_LIB} ${ROCM_HSART_LIB})
+    list(APPEND Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS ${hipblaslt_LIBRARIES})
     list(APPEND Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS
       roc::hipblas hip::hipfft hip::hiprand roc::hipsparse roc::hipsolver)
 

cmake/public/LoadHIP.cmake

@@ -190,6 +190,8 @@ if(HIP_FOUND)
   find_library(ROCM_HIPRTC_LIB hiprtc HINTS ${ROCM_PATH}/lib)
   # roctx is part of roctracer
   find_library(ROCM_ROCTX_LIB roctx64 HINTS ${ROCM_PATH}/lib)
+  # HSA runtime lib
+  find_library(ROCM_HSART_LIB hsa-runtime64 HINTS ${ROCM_PATH}/lib)
 
   if(ROCM_VERSION_DEV VERSION_GREATER_EQUAL "5.7.0")
     # check whether hipblaslt is using its own datatype

torch/lib/libshm/CMakeLists.txt

@@ -60,7 +60,12 @@ if(UNIX AND NOT APPLE)
 endif()
 
 add_executable(torch_shm_manager manager.cpp)
-target_link_libraries(torch_shm_manager PRIVATE shm c10)
+if(BUILD_LIBTORCHLESS)
+  target_link_libraries(torch_shm_manager PRIVATE shm ${C10_LIB})
+else()
+  # we need to link directly to c10 here otherwise we miss symbols
+  target_link_libraries(torch_shm_manager PRIVATE shm c10 ${ROCM_HSART_LIB})
+endif()
 set_target_properties(torch_shm_manager PROPERTIES
   INSTALL_RPATH "${_rpath_portable_origin}/../lib")