git subrepo clone https://github.com/rocm-arch/python-pytorch-rocm

subrepo:
  subdir:   "python-pytorch-rocm"
  merged:   "bc28dc1"
upstream:
  origin:   "https://github.com/rocm-arch/python-pytorch-rocm"
  branch:   "master"
  commit:   "bc28dc1"
git-subrepo:
  version:  "0.4.3"
  origin:   "???"
  commit:   "???"

python-pytorch-rocm/.SRCINFO

@@ -0,0 +1,99 @@
+pkgbase = python-pytorch-rocm
+	pkgdesc = Tensors and Dynamic neural networks in Python with strong GPU acceleration
+	pkgver = 1.7.1
+	pkgrel = 7
+	url = https://pytorch.org
+	arch = x86_64
+	license = BSD
+	makedepends = python
+	makedepends = python-setuptools
+	makedepends = python-yaml
+	makedepends = python-numpy
+	makedepends = cmake
+	makedepends = rocm
+	makedepends = rocm-libs
+	makedepends = miopen
+	makedepends = git
+	makedepends = ninja
+	makedepends = pkgconfig
+	makedepends = doxygen
+	depends = google-glog
+	depends = gflags
+	depends = opencv
+	depends = openmp
+	depends = rccl
+	depends = pybind11
+	depends = python
+	depends = python-yaml
+	depends = libuv
+	depends = python-numpy
+	depends = protobuf
+	depends = ffmpeg
+	depends = python-future
+	depends = qt5-base
+	depends = onednn
+	depends = intel-mkl
+	source = pytorch-1.7.1::git+https://github.com/pytorch/pytorch.git#tag=v1.7.1
+	source = fix_include_system.patch
+	source = use-system-libuv.patch
+	source = use-system-libuv2.patch
+	source = nccl_version.patch
+	source = disable_non_x86_64.patch
+	source = find-hsa-runtime.patch::https://patch-diff.githubusercontent.com/raw/pytorch/pytorch/pull/45550.patch
+	sha256sums = SKIP
+	sha256sums = 83c81ec6a461110da6ae6182529f58100986b068c5182ca62cd53c648b4e4fb0
+	sha256sums = 26b1dd596f1e21a011ee18cab939924483d6c6d4d98e543bf76f5a9312d54d67
+	sha256sums = 7b65c3b209fc39f92ba58a58be6d3da40799f1922910b1171ccd9209eda1f9eb
+	sha256sums = e4a96887b41cbdfd4204ce5f16fcb16a23558d23126331794ab6aa30a66f2e0d
+	sha256sums = d3ef8491718ed7e814fe63e81df2f49862fffbea891d2babbcb464796a1bd680
+	sha256sums = SKIP
+
+pkgname = python-pytorch-rocm
+	pkgdesc = Tensors and Dynamic neural networks in Python with strong GPU acceleration (with ROCM)
+	depends = google-glog
+	depends = gflags
+	depends = opencv
+	depends = openmp
+	depends = rccl
+	depends = pybind11
+	depends = python
+	depends = python-yaml
+	depends = libuv
+	depends = python-numpy
+	depends = protobuf
+	depends = ffmpeg
+	depends = python-future
+	depends = qt5-base
+	depends = onednn
+	depends = intel-mkl
+	depends = rocm
+	depends = rocm-libs
+	depends = miopen
+	provides = python-pytorch
+	conflicts = python-pytorch
+
+pkgname = python-pytorch-opt-rocm
+	pkgdesc = Tensors and Dynamic neural networks in Python with strong GPU acceleration (with ROCM and AVX2 CPU optimizations)
+	depends = google-glog
+	depends = gflags
+	depends = opencv
+	depends = openmp
+	depends = rccl
+	depends = pybind11
+	depends = python
+	depends = python-yaml
+	depends = libuv
+	depends = python-numpy
+	depends = protobuf
+	depends = ffmpeg
+	depends = python-future
+	depends = qt5-base
+	depends = onednn
+	depends = intel-mkl
+	depends = rocm
+	depends = rocm-libs
+	depends = miopen
+	provides = python-pytorch
+	provides = python-pytorch-rocm
+	conflicts = python-pytorch
+

python-pytorch-rocm/.gitrepo

@@ -0,0 +1,12 @@
+; DO NOT EDIT (unless you know what you are doing)
+;
+; This subdirectory is a git "subrepo", and this file is maintained by the
+; git-subrepo command. See https://github.com/git-commands/git-subrepo#readme
+;
+[subrepo]
+	remote = https://github.com/rocm-arch/python-pytorch-rocm
+	branch = master
+	commit = bc28dc180c97c76542995d38276781a7357c784b
+	parent = 91ea9ad6727e316c492bfec38dae8f788432e55e
+	method = merge
+	cmdver = 0.4.3

python-pytorch-rocm/PKGBUILD

@@ -0,0 +1,185 @@
+# Maintainer: acxz <akashpatel2008 at yahoo dot com>
+# Contributor: Sven-Hendrik Haase <svenstaro@gmail.com>
+# Contributor: Stephen Zhang <zsrkmyn at gmail dot com>
+
+pkgbase=python-pytorch-rocm
+
+# Flags for building without/with cpu optimizations
+_build_no_opt=1
+_build_opt=1
+
+pkgname=()
+[ "$_build_no_opt" -eq 1 ] && pkgname+=("python-pytorch-rocm")
+[ "$_build_opt" -eq 1 ]    && pkgname+=("python-pytorch-opt-rocm")
+
+_pkgname="pytorch"
+pkgver=1.7.1
+_pkgver=1.7.1
+pkgrel=7
+pkgdesc="Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+arch=('x86_64')
+url="https://pytorch.org"
+license=('BSD')
+depends=('google-glog' 'gflags' 'opencv' 'openmp' 'rccl' 'pybind11' 'python' 'python-yaml' 'libuv'
+         'python-numpy' 'protobuf' 'ffmpeg' 'python-future' 'qt5-base' 'onednn' 'intel-mkl')
+makedepends=('python' 'python-setuptools' 'python-yaml' 'python-numpy' 'cmake' 'rocm'
+             'rocm-libs' 'miopen' 'git' 'ninja' 'pkgconfig' 'doxygen')
+source=("${_pkgname}-${pkgver}::git+https://github.com/pytorch/pytorch.git#tag=v$_pkgver"
+        fix_include_system.patch
+        use-system-libuv.patch
+        use-system-libuv2.patch
+        nccl_version.patch
+        disable_non_x86_64.patch
+        "find-hsa-runtime.patch::https://patch-diff.githubusercontent.com/raw/pytorch/pytorch/pull/45550.patch")
+sha256sums=('SKIP'
+            '83c81ec6a461110da6ae6182529f58100986b068c5182ca62cd53c648b4e4fb0'
+            '26b1dd596f1e21a011ee18cab939924483d6c6d4d98e543bf76f5a9312d54d67'
+            '7b65c3b209fc39f92ba58a58be6d3da40799f1922910b1171ccd9209eda1f9eb'
+            'e4a96887b41cbdfd4204ce5f16fcb16a23558d23126331794ab6aa30a66f2e0d'
+            'd3ef8491718ed7e814fe63e81df2f49862fffbea891d2babbcb464796a1bd680'
+            'SKIP')
+
+prepare() {
+  cd "${_pkgname}-${pkgver}"
+
+  # This is the lazy way since pytorch has sooo many submodules and they keep
+  # changing them around but we've run into more problems so far doing it the
+  # manual than the lazy way. This lazy way (not explicitly specifying all
+  # submodules) will make building inefficient but for now I'll take it.
+  # It will result in the same package, don't worry.
+  git submodule update --init --recursive
+
+  # https://bugs.archlinux.org/task/64981
+  patch -N torch/utils/cpp_extension.py "${srcdir}"/fix_include_system.patch
+
+  # Use system libuv
+  patch -Np1 -i "${srcdir}"/use-system-libuv.patch
+
+  # FindNCCL patch to export correct nccl version
+  # patch -Np1 -i "${srcdir}"/nccl_version.patch
+
+  # https://github.com/pytorch/pytorch/pull/45550
+  patch -Np1 -i "${srcdir}"/find-hsa-runtime.patch
+
+  # remove local nccl
+  rm -rf third_party/nccl/nccl
+
+  cd ..
+
+  [ "$_build_no_opt" -eq 1 ] && cp -a "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-rocm"
+  [ "$_build_opt" -eq 1 ]    && cp -a "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-opt-rocm"
+
+  export VERBOSE=1
+  export PYTORCH_BUILD_VERSION="${pkgver}"
+  export PYTORCH_BUILD_NUMBER=1
+
+  # Check tools/setup_helpers/cmake.py, setup.py and CMakeLists.txt for a list of flags that can be set via env vars.
+  export USE_MKLDNN=ON
+  export BUILD_CUSTOM_PROTOBUF=ON
+  # export BUILD_SHARED_LIBS=OFF
+  export USE_FFMPEG=ON
+  export USE_GFLAGS=ON
+  export USE_GLOG=ON
+  export BUILD_BINARY=ON
+  export USE_OPENCV=ON
+  export USE_SYSTEM_NCCL=ON
+  # export USE_SYSTEM_LIBS=ON
+  export NCCL_VERSION=$(pkg-config nccl --modversion)
+  export NCCL_VER_CODE=$(sed -n 's/^#define NCCL_VERSION_CODE\s*\(.*\).*/\1/p' /usr/include/nccl.h)
+  export CUDAHOSTCXX=g++
+  export CUDA_HOME=/opt/cuda
+  export CUDNN_LIB_DIR=/usr/lib
+  export CUDNN_INCLUDE_DIR=/usr/include
+  # export TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
+  export TORCH_CUDA_ARCH_LIST="5.2;5.3;6.0;6.1;6.2;7.0;7.0+PTX;7.2;7.2+PTX;7.5;7.5+PTX;8.0;8.0+PTX;8.6;8.6+PTX"
+}
+
+build() {
+  if [ "$_build_no_opt" -eq 1 ]; then
+    echo "Building with rocm and without non-x86-64 optimizations"
+    export USE_CUDA=OFF
+    export USE_ROCM=ON
+    cd "${srcdir}/${_pkgname}-${pkgver}-rocm"
+    patch -Np1 -i "${srcdir}/disable_non_x86_64.patch"
+    echo "add_definitions(-march=x86-64)" >> cmake/MiscCheck.cmake
+
+    # Apply changes needed for ROCm
+    python tools/amd_build/build_amd.py
+
+    # Fix so amdip64 library ending with a dash
+    python setup.py build --cmake-only
+    sed -E -i 's#opt/rocm/hip/lib/libamdhip64\.so\.[0-9.]+-#opt/rocm/hip/lib/libamdhip64.so#' build/build.ninja
+
+    python setup.py build
+  fi
+
+  if [ "$_build_opt" -eq 1 ]; then
+    echo "Building with rocm and with non-x86-64 optimizations"
+    export USE_CUDA=OFF
+    export USE_ROCM=ON
+    cd "${srcdir}/${_pkgname}-${pkgver}-opt-rocm"
+    echo "add_definitions(-march=haswell)" >> cmake/MiscCheck.cmake
+
+    # Apply changes needed for ROCm
+    python tools/amd_build/build_amd.py
+
+    # Fix so amdip64 library ending with a dash
+    python setup.py build --cmake-only
+    sed -E -i 's#opt/rocm/hip/lib/libamdhip64\.so\.[0-9.]+-#opt/rocm/hip/lib/libamdhip64.so#' build/build.ninja
+
+    python setup.py build
+  fi
+}
+
+_package() {
+  # Prevent setup.py from re-running CMake and rebuilding
+  sed -e 's/RUN_BUILD_DEPS = True/RUN_BUILD_DEPS = False/g' -i setup.py
+
+  python setup.py install --root="${pkgdir}"/ --optimize=1 --skip-build
+
+  install -Dm644 LICENSE "${pkgdir}/usr/share/licenses/${pkgname}/LICENSE"
+
+  pytorchpath="usr/lib/python3.9/site-packages/torch"
+  install -d "${pkgdir}/usr/lib"
+
+  # put CMake files in correct place
+  mv "${pkgdir}/${pytorchpath}/share/cmake" "${pkgdir}/usr/lib/cmake"
+
+  # put C++ API in correct place
+  mv "${pkgdir}/${pytorchpath}/include" "${pkgdir}/usr/include"
+  mv "${pkgdir}/${pytorchpath}/lib"/*.so* "${pkgdir}/usr/lib/"
+
+  # clean up duplicates
+  # TODO: move towards direct shared library dependecy of:
+  #   c10, caffe2, libcpuinfo, CUDA RT, gloo, GTest, Intel MKL,
+  #   NVRTC, ONNX, protobuf, libthreadpool, QNNPACK
+  rm -rf "${pkgdir}/usr/include/pybind11"
+
+  # python module is hardcoded to look there at runtime
+  ln -s /usr/include "${pkgdir}/${pytorchpath}/include"
+  find "${pkgdir}"/usr/lib -type f -name "*.so*" -print0 | while read -rd $'\0' _lib; do
+    ln -s ${_lib#"$pkgdir"} "${pkgdir}/${pytorchpath}/lib/"
+  done
+}
+
+package_python-pytorch-rocm() {
+  pkgdesc="Tensors and Dynamic neural networks in Python with strong GPU acceleration (with ROCM)"
+  depends+=(rocm rocm-libs miopen)
+  conflicts=(python-pytorch)
+  provides=(python-pytorch)
+
+  cd "${srcdir}/${_pkgname}-${pkgver}-rocm"
+  _package
+}
+
+package_python-pytorch-opt-rocm() {
+  pkgdesc="Tensors and Dynamic neural networks in Python with strong GPU acceleration (with ROCM and AVX2 CPU optimizations)"
+  depends+=(rocm rocm-libs miopen)
+  conflicts=(python-pytorch)
+  provides=(python-pytorch python-pytorch-rocm)
+
+  cd "${srcdir}/${_pkgname}-${pkgver}-opt-rocm"
+  _package
+}
+
+# vim:set ts=2 sw=2 et:

python-pytorch-rocm/disable_non_x86_64.patch

@@ -0,0 +1,15 @@
+diff --git a/tools/setup_helpers/cmake.py b/tools/setup_helpers/cmake.py
+index d5db749d15..fd54cca6b8 100644
+--- a/tools/setup_helpers/cmake.py
++++ b/tools/setup_helpers/cmake.py
+@@ -295,6 +295,10 @@ class CMake:
+         build_options.update(cmake__options)
+ 
+         CMake.defines(args,
++                      DISABLE_AVX2=1,
++                      DISABLE_AVX512F=1,
++                      DISABLE_FMA4=1,
++                      DISABLE_SSE4=1,
+                       PYTHON_EXECUTABLE=sys.executable,
+                       PYTHON_LIBRARY=cmake_python_library,
+                       PYTHON_INCLUDE_DIR=distutils.sysconfig.get_python_inc(),

python-pytorch-rocm/fix_include_system.patch

@@ -0,0 +1,13 @@
+diff --git a/torch/utils/cpp_extension.py b/torch/utils/cpp_extension.py
+index 7692cad5cd..d6e4a82221 100644
+--- a/torch/utils/cpp_extension.py
++++ b/torch/utils/cpp_extension.py
+@@ -1614,7 +1614,7 @@ def _write_ninja_file_to_build_library(path,
+             common_cflags.append(f'-DPYBIND11_{pname}=\\"{pval}\\"')
+ 
+     common_cflags += [f'-I{include}' for include in user_includes]
+-    common_cflags += [f'-isystem {include}' for include in system_includes]
++    common_cflags += [f'-I{include}' for include in system_includes]
+     common_cflags += ['-D_GLIBCXX_USE_CXX11_ABI=' + str(int(torch._C._GLIBCXX_USE_CXX11_ABI))]
+ 
+     if IS_WINDOWS:

python-pytorch-rocm/nccl_version.patch

@@ -0,0 +1,25 @@
+diff --git a/cmake/Modules/FindNCCL.cmake b/cmake/Modules/FindNCCL.cmake
+index a16c9aca67..51eb54f14c 100644
+--- a/cmake/Modules/FindNCCL.cmake
++++ b/cmake/Modules/FindNCCL.cmake
+@@ -55,9 +55,10 @@ if(NCCL_FOUND)  # obtaining NCCL version and some sanity checks
+   set (OLD_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
+   list (APPEND CMAKE_REQUIRED_INCLUDES ${NCCL_INCLUDE_DIRS})
+   include(CheckCXXSymbolExists)
+-  check_cxx_symbol_exists(NCCL_VERSION_CODE nccl.h NCCL_VERSION_DEFINED)
++  set(NCCL_VERSION_CODE $ENV{NCCL_VER_CODE})
++  set(NCCL_VERSION_DEFINED $ENV{NCCL_VER_CODE})
+ 
+-  if (NCCL_VERSION_DEFINED)
++  if (DEFINED NCCL_VERSION_DEFINED)
+     set(file "${PROJECT_BINARY_DIR}/detect_nccl_version.cc")
+     file(WRITE ${file} "
+       #include <iostream>
+@@ -72,6 +73,7 @@ if(NCCL_FOUND)  # obtaining NCCL version and some sanity checks
+       }
+ ")
+     try_run(NCCL_VERSION_MATCHED compile_result ${PROJECT_BINARY_DIR} ${file}
++          CMAKE_FLAGS -DINCLUDE_DIRECTORIES=/opt/cuda/include
+           RUN_OUTPUT_VARIABLE NCCL_VERSION_FROM_HEADER
+           CMAKE_FLAGS  "-DINCLUDE_DIRECTORIES=${NCCL_INCLUDE_DIRS}"
+           LINK_LIBRARIES ${NCCL_LIBRARIES})

python-pytorch-rocm/use-system-libuv.patch

@@ -0,0 +1,13 @@
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 023bbe9e8d..e25aa0aa7f 100644
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -1308,7 +1308,7 @@ if(USE_DISTRIBUTED AND USE_TENSORPIPE)
+   if(MSVC)
+     message(WARNING "Tensorpipe cannot be used on Windows.")
+   else()
+-    set(TP_BUILD_LIBUV ON CACHE BOOL "" FORCE)
++    set(TP_BUILD_LIBUV OFF CACHE BOOL "" FORCE)
+     set(TP_ENABLE_SHM OFF CACHE BOOL "" FORCE)
+     set(TP_ENABLE_CMA OFF CACHE BOOL "" FORCE)
+     set(TP_STATIC_OR_SHARED STATIC CACHE STRING "" FORCE)

python-pytorch-rocm/use-system-libuv2.patch

@@ -0,0 +1,13 @@
+diff --git a/cmake/pytorch.cmake b/cmake/pytorch.cmake
+index e4d6b9c..48e0669 100644
+--- a/cmake/pytorch.cmake
++++ b/cmake/pytorch.cmake
+@@ -138,7 +138,7 @@ endif()
+ 
+ add_library(tensorpipe ${TENSORPIPE_SRC})
+ 
+-set(TP_BUILD_LIBUV ON)
++set(TP_BUILD_LIBUV OFF)
+ find_package(uv REQUIRED)
+ target_link_libraries(tensorpipe PRIVATE uv::uv)
+