[8.18][ML] Upgrade to PyTorch 2.5 and Boost 1.86 for macOS ARM builds

3rd_party/3rd_party.cmake

@@ -49,7 +49,7 @@ if ("${HOST_SYSTEM_NAME}" STREQUAL "darwin")
   else()
     set(BOOST_ARCH "a64")
   endif()
-  set(BOOST_EXTENSION "mt-${BOOST_ARCH}-1_83.dylib")
+  set(BOOST_EXTENSION "mt-${BOOST_ARCH}-1_86.dylib")
   set(BOOST_LIBRARIES "atomic" "chrono" "date_time" "filesystem" "iostreams" "log" "log_setup" "program_options" "regex" "system" "thread" "unit_test_framework")
   set(XML_LOCATION)
   set(GCC_RT_LOCATION)

build-setup/macos.md

@@ -66,42 +66,42 @@ xcode-select --install
 
 at the command prompt.
 
-### Boost 1.83.0
+### Boost 1.86.0
 
-Download version 1.83.0 of Boost from <https://boostorg.jfrog.io/artifactory/main/release/1.83.0/source/boost_1_83_0.tar.bz2>. You must get this exact version, as the Machine Learning build system requires it.
+Download version 1.86.0 of Boost from <https://boostorg.jfrog.io/artifactory/main/release/1.86.0/source/boost_1_86_0.tar.bz2>. You must get this exact version, as the Machine Learning build system requires it.
 
 Assuming you chose the `.bz2` version, extract it to a temporary directory:
 
 ```
-bzip2 -cd boost_1_83_0.tar.bz2 | tar xvf -
+bzip2 -cd boost_1_86_0.tar.bz2 | tar xvf -
 ```
 
-In the resulting `boost_1_83_0` directory, run:
+In the resulting `boost_1_86_0` directory, run:
 
 ```
 ./bootstrap.sh --with-toolset=clang --without-libraries=context --without-libraries=coroutine --without-libraries=graph_parallel --without-libraries=mpi --without-libraries=python --without-icu
 ```
 
 This should build the `b2` program, which in turn is used to build Boost.
 
-Edit `boost/unordered/detail/prime_fmod.hpp` and change line 134 from
+Edit `boost/unordered/detail/prime_fmod.hpp` and change line 37 from
 
 ```
-    (13ul)(29ul)(53ul)(97ul)(193ul)(389ul)(769ul)(1543ul)(3079ul)(6151ul)(       \
+    constexpr static std::size_t const sizes[] = {13ul, 29ul, 53ul, 97ul,
 ```
 
 to:
 
 ```
-    (3ul)(13ul)(29ul)(53ul)(97ul)(193ul)(389ul)(769ul)(1543ul)(3079ul)(6151ul)(       \
+    constexpr static std::size_t const sizes[] = {3ul, 13ul, 29ul, 53ul, 97ul,
 ```
 
 
 To complete the build, type:
 
 ```
 ./b2 -j8 --layout=versioned --disable-icu cxxflags="-std=c++17 -stdlib=libc++ $SSEFLAGS" linkflags="-std=c++17 -stdlib=libc++ -Wl,-headerpad_max_install_names" optimization=speed inlining=full define=BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS define=BOOST_LOG_WITHOUT_DEBUG_OUTPUT define=BOOST_LOG_WITHOUT_EVENT_LOG define=BOOST_LOG_WITHOUT_SYSLOG define=BOOST_LOG_WITHOUT_IPC
-sudo ./b2 install --layout=versioned --disable-icu cxxflags="-std=c++17 -stdlib=libc++ $SSEFLAGS" linkflags="-std=c++17 -stdlib=libc++ -Wl,-headerpad_max_install_names" optimization=speed inlining=full define=BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS define=BOOST_LOG_WITHOUT_DEBUG_OUTPUT define=BOOST_LOG_WITHOUT_EVENT_LOG define=BOOST_LOG_WITHOUT_SYSLOG define=BOOST_LOG_WITHOUT_IPC
+sudo ./b2 -j8 install --layout=versioned --disable-icu cxxflags="-std=c++17 -stdlib=libc++ $SSEFLAGS" linkflags="-std=c++17 -stdlib=libc++ -Wl,-headerpad_max_install_names" optimization=speed inlining=full define=BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS define=BOOST_LOG_WITHOUT_DEBUG_OUTPUT define=BOOST_LOG_WITHOUT_EVENT_LOG define=BOOST_LOG_WITHOUT_SYSLOG define=BOOST_LOG_WITHOUT_IPC
 ```
 
 to install the Boost headers and libraries.
@@ -127,7 +127,7 @@ Download the graphical installer for Python 3.10.9 from <https://www.python.org/
 
 Install using all the default options.  When the installer completes a Finder window pops up.  Double click the `Install Certificates.command` file in this folder to install the SSL certificates Python needs.
 
-### PyTorch 2.3.1
+### PyTorch 2.5.0
 
 PyTorch requires that certain Python modules are installed.  To install them:
 
@@ -138,7 +138,7 @@ sudo /Library/Frameworks/Python.framework/Versions/3.10/bin/pip3.10 install nump
 Then obtain the PyTorch code:
 
 ```
-git clone --depth=1 --branch=v2.3.1 https://github.com/pytorch/pytorch.git
+git clone --depth=1 --branch=v2.5.0 https://github.com/pytorch/pytorch.git
 cd pytorch
 git submodule sync
 git submodule update --init --recursive
@@ -170,6 +170,38 @@ inline void to_bytes(bytestring& bytes, const unsigned long arg) {
 
 at around line 189. This is necessary to resolve a template specialization issue.
 
+Depending on your version of macOS it may also be necessary to follow these additional steps:
+
+* Edit `aten/src/ATen/mps/MPSDevice.mm` and change
+  ```
+  return MTLLanguageVersion3_0;
+  ```
+  to
+  ```
+  return MTLLanguageVersion2_0;
+  ```
+* Edit `aten/src/ATen/native/mps/operations/Indexing.mm` and change
+  ```
+  indexABContents[idx] =
+      getMTLBufferStorage(indexTensor).gpuAddress + (indexTensor.storage_offset() * indexTensor.element_size());
+  ```
+  to
+  ```
+  //indexABContents[idx] =
+  //    getMTLBufferStorage(indexTensor).gpuAddress + (indexTensor.storage_offset() * indexTensor.element_size());
+  ```
+* Edit `third_party/onnx/CMakelists.txt` and change
+  ```
+  # find_package(Python ${PY_VERSION} COMPONENTS Interpreter Development REQUIRED)
+  find_package(PythonInterp ${PY_VERSION} REQUIRED)
+  ```
+  to
+  ```
+  set(PY_VERSION 3.10)
+  find_package(Python3 ${PY_VERSION} COMPONENTS Interpreter Development REQUIRED)
+  #find_package(PythonInterp ${PY_VERSION} REQUIRED)
+  ```
+
 For compilation on `macOS Mojave (10.14)`
 
 * Edit `torch/csrc/jit/api/module.h` and change
@@ -216,12 +248,11 @@ export BUILD_TEST=OFF
 export BUILD_CAFFE2=OFF
 export USE_NUMPY=OFF
 export USE_DISTRIBUTED=OFF
-[ $(uname -m) != x86_64 ] && export DNNL_TARGET_ARCH=AARCH64
+export DNNL_TARGET_ARCH=AARCH64
 export USE_MKLDNN=ON
 export USE_QNNPACK=OFF
 export USE_PYTORCH_QNNPACK=OFF
-[ $(uname -m) = x86_64 ] && export USE_XNNPACK=OFF
-export PYTORCH_BUILD_VERSION=2.3.1
+export PYTORCH_BUILD_VERSION=2.5.0
 export PYTORCH_BUILD_NUMBER=1
 /Library/Frameworks/Python.framework/Versions/3.10/bin/python3.10 setup.py install
 ```

cmake/variables.cmake

@@ -243,9 +243,14 @@ set(Boost_USE_STATIC_LIBS OFF)
 set(Boost_USE_DEBUG_RUNTIME OFF)
 set(Boost_COMPILER "${ML_BOOST_COMPILER_VER}")
 
-set(Boost_VERSION 1.83.0)
+set (Boost_VERSION 1.83.0)
+if (CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_HOST_SYSTEM_PROCESSOR STREQUAL "arm64")
+  message(WARNING "Using Boost 1.86.0 on macOS.")
+  set (Boost_VERSION 1.86.0)
+endif()
+
 if (CMAKE_SYSTEM_NAME STREQUAL "Windows")
-  message(WARNING "Using Boost 1.86.0 on Windows only.")
+  message(WARNING "Using Boost 1.86.0 on Windows.")
   set(Boost_VERSION 1.86.0)
 endif()
 find_package(Boost ${Boost_VERSION} EXACT REQUIRED COMPONENTS iostreams filesystem program_options regex date_time log log_setup thread unit_test_framework)

dev-tools/download_macos_deps.sh

@@ -28,7 +28,7 @@ case `uname -m` in
         ARCHIVE=local-x86_64-apple-macosx12.0-1.tar.bz2
         ;;
     arm64)
-        ARCHIVE=local-arm64-apple-macosx11.1-10.tar.bz2
+        ARCHIVE=local-arm64-apple-macosx11.1-11.tar.bz2
         ;;
 
     *)

docs/CHANGELOG.asciidoc

@@ -32,8 +32,8 @@
 
 === Enhancements
 
-* Update the PyTorch library to version 2.5.0. (See {ml-pull}2783[#2783].)
-* Upgrade Boost libraries to version 1.86. (See {ml-pull}2780[#2780].)
+* Update the PyTorch library to version 2.5.0. (See {ml-pull}2783[#2783], {ml-pull}2778[#2778].)
+* Upgrade Boost libraries to version 1.86. (See {ml-pull}2780[#2780], {ml-pull}2779[#2779].)
 
 == {es} version 8.16.0
 

lib/core/unittest/CMemoryUsageTest.cc

@@ -1292,7 +1292,7 @@ BOOST_AUTO_TEST_CASE(testSmallVector) {
     // off of internal storage it would NEVER go back to internal storage.
     // Arguably that was a bug, and this assertion started failing after
     // upgrading Boost to 1.86.0, meaning that boost::small_vector has been improved.
-    BOOST_TEST_REQUIRE(extraMem >= 0); // Change this to "==" when all platforms have been upgraded to Boost 1.86.0
+    BOOST_TEST_REQUIRE(extraMem >= 0); // Change to `==` once upgraded to Boost 1.86 on all platforms
 }
 
 BOOST_AUTO_TEST_CASE(testAlignedVector) {