diff --git a/packages/py-torch/cusparseGetErrorString.patch b/packages/py-torch/cusparseGetErrorString.patch
new file mode 100644
index 0000000000000000000000000000000000000000..9cb136b3f88faee0631eaac14c2915664e30dc6c
--- /dev/null
+++ b/packages/py-torch/cusparseGetErrorString.patch
@@ -0,0 +1,53 @@
+diff --git a/aten/src/ATen/native/sparse/cuda/SparseCUDABlas.cu b/aten/src/ATen/native/sparse/cuda/SparseCUDABlas.cu
+index 1cee04c200..f46003d9a9 100644
+--- a/aten/src/ATen/native/sparse/cuda/SparseCUDABlas.cu
++++ b/aten/src/ATen/native/sparse/cuda/SparseCUDABlas.cu
+@@ -10,48 +10,6 @@
+ namespace at { namespace native { namespace sparse { namespace cuda {
+
+
+-std::string cusparseGetErrorString(cusparseStatus_t status) {
+- switch(status)
+- {
+- case CUSPARSE_STATUS_SUCCESS:
+- return "success";
+-
+- case CUSPARSE_STATUS_NOT_INITIALIZED:
+- return "library not initialized";
+-
+- case CUSPARSE_STATUS_ALLOC_FAILED:
+- return "resource allocation failed";
+-
+- case CUSPARSE_STATUS_INVALID_VALUE:
+- return "an invalid numeric value was used as an argument";
+-
+- case CUSPARSE_STATUS_ARCH_MISMATCH:
+- return "an absent device architectural feature is required";
+-
+- case CUSPARSE_STATUS_MAPPING_ERROR:
+- return "an access to GPU memory space failed";
+-
+- case CUSPARSE_STATUS_EXECUTION_FAILED:
+- return "the GPU program failed to execute";
+-
+- case CUSPARSE_STATUS_INTERNAL_ERROR:
+- return "an internal operation failed";
+-
+- case CUSPARSE_STATUS_MATRIX_TYPE_NOT_SUPPORTED:
+- return "the matrix type is not supported by this function";
+-
+- case CUSPARSE_STATUS_ZERO_PIVOT:
+- return "an entry of the matrix is either structural zero or numerical zero (singular block)";
+-
+- default:
+- {
+- std::ostringstream oss;
+- oss << "unknown error " << static_cast<int64_t>(status);
+- return oss.str();
+- }
+- }
+-}
+-
+ inline void CUSPARSE_CHECK(cusparseStatus_t status)
+ {
+ if (status != CUSPARSE_STATUS_SUCCESS) {
diff --git a/packages/py-torch/detect_omp_of_fujitsu_compiler.patch b/packages/py-torch/detect_omp_of_fujitsu_compiler.patch
new file mode 100644
index 0000000000000000000000000000000000000000..519d66869d578ea4a59c4e7f626569baade6837a
--- /dev/null
+++ b/packages/py-torch/detect_omp_of_fujitsu_compiler.patch
@@ -0,0 +1,20 @@
+--- pytorch/cmake/Modules/FindOpenMP.cmake.org 2020-05-26 17:43:53.000000000 +0900
++++ pytorch/cmake/Modules/FindOpenMP.cmake 2020-05-26 17:46:37.000000000 +0900
+@@ -84,7 +84,7 @@
+ unset(OpenMP_FLAG_CANDIDATES)
+
+ set(OMP_FLAG_GNU "-fopenmp")
+- set(OMP_FLAG_Clang "-fopenmp=libomp" "-fopenmp=libiomp5" "-fopenmp")
++ set(OMP_FLAG_Clang "-fopenmp" "-fopenmp=libomp" "-fopenmp=libiomp5")
+
+ # AppleClang may need a header file, search for omp.h with hints to brew
+ # default include dir
+@@ -245,7 +245,7 @@
+ set(OpenMP_libomp_LIBRARY "${MKL_OPENMP_LIBRARY}" CACHE STRING "libomp location for OpenMP")
+ else()
+ find_library(OpenMP_libomp_LIBRARY
+- NAMES omp gomp iomp5
++ NAMES fjomp omp gomp iomp5
+ HINTS ${CMAKE_${LANG}_IMPLICIT_LINK_DIRECTORIES}
+ DOC "libomp location for OpenMP"
+ )
diff --git a/packages/py-torch/package.py b/packages/py-torch/package.py
new file mode 100644
index 0000000000000000000000000000000000000000..de0c36bea5240150e4539069c405233b7c685a9b
--- /dev/null
+++ b/packages/py-torch/package.py
@@ -0,0 +1,407 @@
+# Copyright 2013-2022 Lawrence Livermore National Security, LLC and other
+# Spack Project Developers. See the top-level COPYRIGHT file for details.
+#
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+import os
+import sys
+
+from spack import *
+
+
+class PyTorch(PythonPackage, CudaPackage):
+ """Tensors and Dynamic neural networks in Python
+ with strong GPU acceleration."""
+
+ homepage = "https://pytorch.org/"
+ git = "https://github.com/pytorch/pytorch.git"
+
+ maintainers = ['adamjstewart']
+
+ # Exact set of modules is version- and variant-specific, just attempt to import the
+ # core libraries to ensure that the package was successfully installed.
+ import_modules = ['torch', 'torch.autograd', 'torch.nn', 'torch.utils']
+
+ version('master', branch='master', submodules=True)
+ version('1.10.2', tag='v1.10.2', submodules=True)
+ version('1.10.1', tag='v1.10.1', submodules=True)
+ version('1.10.0', tag='v1.10.0', submodules=True)
+ version('1.9.1', tag='v1.9.1', submodules=True)
+ version('1.9.0', tag='v1.9.0', submodules=True)
+ version('1.8.2', tag='v1.8.2', submodules=True)
+ version('1.8.1', tag='v1.8.1', submodules=True)
+ version('1.8.0', tag='v1.8.0', submodules=True)
+ version('1.7.1', tag='v1.7.1', submodules=True)
+ version('1.7.0', tag='v1.7.0', submodules=True)
+ version('1.6.0', tag='v1.6.0', submodules=True)
+ version('1.5.1', tag='v1.5.1', submodules=True)
+ version('1.5.0', tag='v1.5.0', submodules=True)
+ version('1.4.1', tag='v1.4.1', submodules=True)
+ version('1.3.1', tag='v1.3.1', submodules=True)
+ version('1.3.0', tag='v1.3.0', submodules=True)
+ version('1.2.0', tag='v1.2.0', submodules=True)
+ version('1.1.0', tag='v1.1.0', submodules=True)
+ version('1.0.1', tag='v1.0.1', submodules=True, deprecated=True)
+ version('1.0.0', tag='v1.0.0', submodules=True, deprecated=True)
+
+ is_darwin = sys.platform == 'darwin'
+
+ # All options are defined in CMakeLists.txt.
+ # Some are listed in setup.py, but not all.
+ variant('caffe2', default=True, description='Build Caffe2', when='@1.7:')
+ variant('test', default=False, description='Build C++ test binaries')
+ variant('cuda', default=not is_darwin, description='Use CUDA')
+ variant('rocm', default=False, description='Use ROCm')
+ variant('cudnn', default=not is_darwin, description='Use cuDNN', when='+cuda')
+ variant('fbgemm', default=True, description='Use FBGEMM (quantized 8-bit server operators)')
+ variant('kineto', default=True, description='Use Kineto profiling library', when='@1.8:')
+ variant('magma', default=not is_darwin, description='Use MAGMA', when='+cuda')
+ variant('metal', default=is_darwin, description='Use Metal for Caffe2 iOS build')
+ variant('nccl', default=True, description='Use NCCL', when='+cuda platform=linux')
+ variant('nccl', default=True, description='Use NCCL', when='+cuda platform=cray')
+ variant('nccl', default=True, description='Use NCCL', when='+rocm platform=linux')
+ variant('nccl', default=True, description='Use NCCL', when='+rocm platform=cray')
+ variant('nnpack', default=True, description='Use NNPACK')
+ variant('numa', default=True, description='Use NUMA', when='platform=linux')
+ variant('numa', default=True, description='Use NUMA', when='platform=cray')
+ variant('numpy', default=True, description='Use NumPy')
+ variant('openmp', default=True, description='Use OpenMP for parallel code')
+ variant('qnnpack', default=True, description='Use QNNPACK (quantized 8-bit operators)')
+ variant('valgrind', default=True, description='Use Valgrind', when='@1.8: platform=linux')
+ variant('valgrind', default=True, description='Use Valgrind', when='@1.8: platform=cray')
+ variant('xnnpack', default=True, description='Use XNNPACK', when='@1.5:')
+ variant('mkldnn', default=True, description='Use MKLDNN')
+ variant('distributed', default=not is_darwin, description='Use distributed')
+ variant('mpi', default=not is_darwin, description='Use MPI for Caffe2', when='+distributed')
+ variant('gloo', default=not is_darwin, description='Use Gloo', when='+distributed')
+ variant('tensorpipe', default=not is_darwin, description='Use TensorPipe', when='@1.6: +distributed')
+ variant('onnx_ml', default=True, description='Enable traditional ONNX ML API', when='@1.5:')
+ variant('breakpad', default=True, description='Enable breakpad crash dump library', when='@1.9:')
+
+ conflicts('+cuda+rocm')
+ conflicts('+breakpad', when='target=ppc64:')
+ conflicts('+breakpad', when='target=ppc64le:')
+
+ conflicts('cuda_arch=none', when='+cuda',
+ msg='Must specify CUDA compute capabilities of your GPU, see '
+ 'https://developer.nvidia.com/cuda-gpus')
+
+ # Required dependencies
+ depends_on('cmake@3.5:', type='build')
+ # Use Ninja generator to speed up build times, automatically used if found
+ depends_on('ninja@1.5:', when='@1.1:', type='build')
+ # See python_min_version in setup.py
+ depends_on('python@3.6.2:', when='@1.7.1:', type=('build', 'link', 'run'))
+ depends_on('python@3.6.1:', when='@1.6:1.7.0', type=('build', 'link', 'run'))
+ depends_on('python@3.5:', when='@1.5', type=('build', 'link', 'run'))
+ depends_on('python@2.7:2,3.5:', when='@1.4', type=('build', 'link', 'run'))
+ depends_on('python@2.7:2,3.5:3.7', when='@:1.3', type=('build', 'link', 'run'))
+ depends_on('py-setuptools', type=('build', 'run'))
+ depends_on('py-future', when='@1.5:', type=('build', 'run'))
+ depends_on('py-future', when='@1.1: ^python@:2', type=('build', 'run'))
+ depends_on('py-pyyaml', type=('build', 'run'))
+ depends_on('py-typing', when='^python@:3.4', type=('build', 'run'))
+ depends_on('py-pybind11@2.6.2:', when='@1.8:', type=('build', 'link', 'run'))
+ depends_on('py-pybind11@2.3.0', when='@1.1:1.7', type=('build', 'link', 'run'))
+ depends_on('py-pybind11@2.2.4', when='@:1.0', type=('build', 'link', 'run'))
+ depends_on('py-dataclasses', when='@1.7: ^python@3.6', type=('build', 'run'))
+ depends_on('py-tqdm', type='run')
+ # https://github.com/onnx/onnx#prerequisites
+ depends_on('py-numpy@1.16.6:', type=('build', 'run'))
+ depends_on('py-protobuf@3.12.2:', when='@1.10:', type=('build', 'run'))
+ depends_on('py-protobuf@:3.14', when='@:1.9', type=('build', 'run'))
+ depends_on('protobuf@3.12.2:', when='@1.10:')
+ depends_on('protobuf@:3.14', when='@:1.9')
+ depends_on('py-typing-extensions@3.6.2.1:', when='@1.7:', type=('build', 'run'))
+ depends_on('blas')
+ depends_on('lapack')
+ depends_on('eigen')
+ # https://github.com/pytorch/pytorch/issues/60329
+ # depends_on('cpuinfo@2020-12-17', when='@1.8:')
+ # depends_on('cpuinfo@2020-06-11', when='@1.6:1.7')
+ # https://github.com/shibatch/sleef/issues/427
+ # depends_on('sleef@3.5.1_2020-12-22', when='@1.8:')
+ # https://github.com/pytorch/pytorch/issues/60334
+ # depends_on('sleef@3.4.0_2019-07-30', when='@1.6:1.7')
+ # https://github.com/Maratyszcza/FP16/issues/18
+ # depends_on('fp16@2020-05-14', when='@1.6:')
+ depends_on('pthreadpool@2021-04-13', when='@1.9:')
+ depends_on('pthreadpool@2020-10-05', when='@1.8')
+ depends_on('pthreadpool@2020-06-15', when='@1.6:1.7')
+ depends_on('psimd@2020-05-17', when='@1.6:')
+ depends_on('fxdiv@2020-04-17', when='@1.6:')
+ depends_on('benchmark', when='@1.6:+test')
+
+ # Optional dependencies
+ # https://discuss.pytorch.org/t/compiling-1-10-1-from-source-with-gcc-11-and-cuda-11-5/140971
+ depends_on('cuda@9.2:', when='@1.11:+cuda', type=('build', 'link', 'run'))
+ depends_on('cuda@9.2:11.4', when='@1.6:+cuda', type=('build', 'link', 'run'))
+ depends_on('cuda@9:11.4', when='@1.1:+cuda', type=('build', 'link', 'run'))
+ depends_on('cuda@7.5:11.4', when='+cuda', type=('build', 'link', 'run'))
+ depends_on('cudnn@6:7', when='@:1.0+cudnn')
+ depends_on('cudnn@7.0:7', when='@1.1:1.5+cudnn')
+ depends_on('cudnn@7:', when='@1.6:+cudnn')
+ depends_on('magma', when='+magma')
+ depends_on('nccl', when='+nccl')
+ depends_on('numactl', when='+numa')
+ depends_on('llvm-openmp', when='%apple-clang +openmp')
+ depends_on('valgrind', when='+valgrind')
+ # https://github.com/pytorch/pytorch/issues/60332
+ # depends_on('xnnpack@2021-02-22', when='@1.8:+xnnpack')
+ # depends_on('xnnpack@2020-03-23', when='@1.6:1.7+xnnpack')
+ depends_on('mpi', when='+mpi')
+ # https://github.com/pytorch/pytorch/issues/60270
+ # depends_on('gloo@2021-05-04', when='@1.9:+gloo')
+ # depends_on('gloo@2020-09-18', when='@1.7:1.8+gloo')
+ # depends_on('gloo@2020-03-17', when='@1.6+gloo')
+ # https://github.com/pytorch/pytorch/issues/60331
+ # depends_on('onnx@1.8.0_2020-11-03', when='@1.8:+onnx_ml')
+ # depends_on('onnx@1.7.0_2020-05-31', when='@1.6:1.7+onnx_ml')
+ depends_on('mkl', when='+mkldnn')
+
+ # Test dependencies
+ depends_on('py-hypothesis', type='test')
+ depends_on('py-six', type='test')
+ depends_on('py-psutil', type='test')
+
+ # Fix BLAS being overridden by MKL
+ # https://github.com/pytorch/pytorch/issues/60328
+ patch('https://patch-diff.githubusercontent.com/raw/pytorch/pytorch/pull/59220.patch',
+ sha256='e37afffe45cf7594c22050109942370e49983ad772d12ebccf508377dc9dcfc9',
+ when='@1.2:')
+
+ # Fixes build on older systems with glibc <2.12
+ patch('https://patch-diff.githubusercontent.com/raw/pytorch/pytorch/pull/55063.patch',
+ sha256='e17eaa42f5d7c18bf0d7c37d7b0910127a01ad53fdce3e226a92893356a70395',
+ when='@1.1:1.8.1')
+
+ # Fixes CMake configuration error when XNNPACK is disabled
+ # https://github.com/pytorch/pytorch/pull/35607
+ # https://github.com/pytorch/pytorch/pull/37865
+ patch('xnnpack.patch', when='@1.5')
+
+ # Fixes build error when ROCm is enabled for pytorch-1.5 release
+ patch('rocm.patch', when='@1.5+rocm')
+
+ # Fixes fatal error: sleef.h: No such file or directory
+ # https://github.com/pytorch/pytorch/pull/35359
+ # https://github.com/pytorch/pytorch/issues/26555
+ # patch('sleef.patch', when='@:1.5')
+
+ # Fixes compilation with Clang 9.0.0 and Apple Clang 11.0.3
+ # https://github.com/pytorch/pytorch/pull/37086
+ patch('https://github.com/pytorch/pytorch/commit/e921cd222a8fbeabf5a3e74e83e0d8dfb01aa8b5.patch',
+ sha256='17561b16cd2db22f10c0fe1fdcb428aecb0ac3964ba022a41343a6bb8cba7049',
+ when='@1.1:1.5')
+
+ # Removes duplicate definition of getCusparseErrorString
+ # https://github.com/pytorch/pytorch/issues/32083
+ patch('cusparseGetErrorString.patch', when='@:1.0^cuda@10.1.243:')
+
+ # Fixes 'FindOpenMP.cmake'
+ # to detect openmp settings used by Fujitsu compiler.
+ patch('detect_omp_of_fujitsu_compiler.patch', when='%fj')
+
+ # Fix compilation of +distributed~tensorpipe
+ # https://github.com/pytorch/pytorch/issues/68002
+ patch('https://github.com/pytorch/pytorch/commit/c075f0f633fa0136e68f0a455b5b74d7b500865c.patch',
+ sha256='e69e41b5c171bfb00d1b5d4ee55dd5e4c8975483230274af4ab461acd37e40b8', when='@1.10.0+distributed~tensorpipe')
+
+ @property
+ def libs(self):
+ # TODO: why doesn't `python_platlib` work here?
+ root = join_path(
+ self.prefix, self.spec['python'].package.platlib, 'torch', 'lib'
+ )
+ return find_libraries('libtorch', root)
+
+ @property
+ def headers(self):
+ # TODO: why doesn't `python_platlib` work here?
+ root = join_path(
+ self.prefix, self.spec['python'].package.platlib, 'torch', 'include'
+ )
+ headers = find_all_headers(root)
+ headers.directories = [root]
+ return headers
+
+ @when('@1.5.0:')
+ def patch(self):
+ # https://github.com/pytorch/pytorch/issues/52208
+ filter_file('torch_global_deps PROPERTIES LINKER_LANGUAGE C',
+ 'torch_global_deps PROPERTIES LINKER_LANGUAGE CXX',
+ 'caffe2/CMakeLists.txt')
+
+ def setup_build_environment(self, env):
+ """Set environment variables used to control the build.
+
+ PyTorch's ``setup.py`` is a thin wrapper around ``cmake``.
+ In ``tools/setup_helpers/cmake.py``, you can see that all
+ environment variables that start with ``BUILD_``, ``USE_``,
+ or ``CMAKE_``, plus a few more explicitly specified variable
+ names, are passed directly to the ``cmake`` call. Therefore,
+ most flags defined in ``CMakeLists.txt`` can be specified as
+ environment variables.
+ """
+ def enable_or_disable(variant, keyword='USE', var=None, newer=False):
+ """Set environment variable to enable or disable support for a
+ particular variant.
+
+ Parameters:
+ variant (str): the variant to check
+ keyword (str): the prefix to use for enabling/disabling
+ var (str): CMake variable to set. Defaults to variant.upper()
+ newer (bool): newer variants that never used NO_*
+ """
+ if var is None:
+ var = variant.upper()
+
+ # Version 1.1.0 switched from NO_* to USE_* or BUILD_*
+ # But some newer variants have always used USE_* or BUILD_*
+ if self.spec.satisfies('@1.1:') or newer:
+ if '+' + variant in self.spec:
+ env.set(keyword + '_' + var, 'ON')
+ elif '~' + variant in self.spec:
+ env.set(keyword + '_' + var, 'OFF')
+ else:
+ if '+' + variant in self.spec:
+ env.unset('NO_' + var)
+ elif '~' + variant in self.spec:
+ env.set('NO_' + var, 'ON')
+
+ # Build in parallel to speed up build times
+ env.set('MAX_JOBS', make_jobs)
+
+ # Spack logs have trouble handling colored output
+ env.set('COLORIZE_OUTPUT', 'OFF')
+
+ enable_or_disable('test', keyword='BUILD')
+ enable_or_disable('caffe2', keyword='BUILD')
+
+ enable_or_disable('cuda')
+ if '+cuda' in self.spec:
+ # cmake/public/cuda.cmake
+ # cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
+ env.unset('CUDA_ROOT')
+ torch_cuda_arch = ';'.join('{0:.1f}'.format(float(i) / 10.0) for i
+ in
+ self.spec.variants['cuda_arch'].value)
+ env.set('TORCH_CUDA_ARCH_LIST', torch_cuda_arch)
+
+ enable_or_disable('rocm')
+
+ enable_or_disable('cudnn')
+ if '+cudnn' in self.spec:
+ # cmake/Modules_CUDA_fix/FindCUDNN.cmake
+ env.set('CUDNN_INCLUDE_DIR', self.spec['cudnn'].prefix.include)
+ env.set('CUDNN_LIBRARY', self.spec['cudnn'].libs[0])
+
+ enable_or_disable('fbgemm')
+ enable_or_disable('kineto')
+ enable_or_disable('magma')
+ enable_or_disable('metal')
+ enable_or_disable('breakpad')
+
+ enable_or_disable('nccl')
+ if '+nccl' in self.spec:
+ env.set('NCCL_LIB_DIR', self.spec['nccl'].libs.directories[0])
+ env.set('NCCL_INCLUDE_DIR', self.spec['nccl'].prefix.include)
+
+ # cmake/External/nnpack.cmake
+ enable_or_disable('nnpack')
+
+ enable_or_disable('numa')
+ if '+numa' in self.spec:
+ # cmake/Modules/FindNuma.cmake
+ env.set('NUMA_ROOT_DIR', self.spec['numactl'].prefix)
+
+ # cmake/Modules/FindNumPy.cmake
+ enable_or_disable('numpy')
+ # cmake/Modules/FindOpenMP.cmake
+ enable_or_disable('openmp', newer=True)
+ enable_or_disable('qnnpack')
+ enable_or_disable('qnnpack', var='PYTORCH_QNNPACK')
+ enable_or_disable('valgrind')
+ enable_or_disable('xnnpack')
+ enable_or_disable('mkldnn')
+ enable_or_disable('distributed')
+ enable_or_disable('mpi')
+ # cmake/Modules/FindGloo.cmake
+ enable_or_disable('gloo', newer=True)
+ enable_or_disable('tensorpipe')
+
+ if '+onnx_ml' in self.spec:
+ env.set('ONNX_ML', 'ON')
+ elif '~onnx_ml' in self.spec:
+ env.set('ONNX_ML', 'OFF')
+
+ if not self.spec.satisfies('@master'):
+ env.set('PYTORCH_BUILD_VERSION', self.version)
+ env.set('PYTORCH_BUILD_NUMBER', 0)
+
+ # BLAS to be used by Caffe2
+ # Options defined in cmake/Dependencies.cmake and cmake/Modules/FindBLAS.cmake
+ if self.spec['blas'].name == 'atlas':
+ env.set('BLAS', 'ATLAS')
+ env.set('WITH_BLAS', 'atlas')
+ elif self.spec['blas'].name in ['blis', 'amdblis']:
+ env.set('BLAS', 'BLIS')
+ env.set('WITH_BLAS', 'blis')
+ elif self.spec['blas'].name == 'eigen':
+ env.set('BLAS', 'Eigen')
+ elif self.spec['lapack'].name in ['libflame', 'amdlibflame']:
+ env.set('BLAS', 'FLAME')
+ env.set('WITH_BLAS', 'FLAME')
+ elif self.spec['blas'].name in [
+ 'intel-mkl', 'intel-parallel-studio', 'intel-oneapi-mkl']:
+ env.set('BLAS', 'MKL')
+ env.set('WITH_BLAS', 'mkl')
+ elif self.spec['blas'].name == 'openblas':
+ env.set('BLAS', 'OpenBLAS')
+ env.set('WITH_BLAS', 'open')
+ elif self.spec['blas'].name == 'veclibfort':
+ env.set('BLAS', 'vecLib')
+ env.set('WITH_BLAS', 'veclib')
+ else:
+ env.set('BLAS', 'Generic')
+ env.set('WITH_BLAS', 'generic')
+
+ # Don't use vendored third-party libraries when possible
+ env.set('BUILD_CUSTOM_PROTOBUF', 'OFF')
+ env.set('USE_SYSTEM_NCCL', 'ON')
+ env.set('USE_SYSTEM_EIGEN_INSTALL', 'ON')
+ env.set('pybind11_DIR', self.spec['py-pybind11'].prefix)
+ env.set('pybind11_INCLUDE_DIR',
+ self.spec['py-pybind11'].prefix.include)
+ if self.spec.satisfies('@1.10:'):
+ env.set('USE_SYSTEM_PYBIND11', 'ON')
+ # https://github.com/pytorch/pytorch/issues/60334
+ # if self.spec.satisfies('@1.8:'):
+ # env.set('USE_SYSTEM_SLEEF', 'ON')
+ if self.spec.satisfies('@1.6:'):
+ # env.set('USE_SYSTEM_LIBS', 'ON')
+ # https://github.com/pytorch/pytorch/issues/60329
+ # env.set('USE_SYSTEM_CPUINFO', 'ON')
+ # https://github.com/pytorch/pytorch/issues/60270
+ # env.set('USE_SYSTEM_GLOO', 'ON')
+ # https://github.com/Maratyszcza/FP16/issues/18
+ # env.set('USE_SYSTEM_FP16', 'ON')
+ env.set('USE_SYSTEM_PTHREADPOOL', 'ON')
+ env.set('USE_SYSTEM_PSIMD', 'ON')
+ env.set('USE_SYSTEM_FXDIV', 'ON')
+ env.set('USE_SYSTEM_BENCHMARK', 'ON')
+ # https://github.com/pytorch/pytorch/issues/60331
+ # env.set('USE_SYSTEM_ONNX', 'ON')
+ # https://github.com/pytorch/pytorch/issues/60332
+ # env.set('USE_SYSTEM_XNNPACK', 'ON')
+
+ @run_before('install')
+ def build_amd(self):
+ if '+rocm' in self.spec:
+ python(os.path.join('tools', 'amd_build', 'build_amd.py'))
+
+ @run_after('install')
+ @on_package_attributes(run_tests=True)
+ def install_test(self):
+ with working_dir('test'):
+ python('run_test.py')
diff --git a/packages/py-torch/rocm.patch b/packages/py-torch/rocm.patch
new file mode 100644
index 0000000000000000000000000000000000000000..b50cc7e1598a23f41e1e1a73e6672e6a4d132b6a
--- /dev/null
+++ b/packages/py-torch/rocm.patch
@@ -0,0 +1,98 @@
+diff --git a/aten/src/ATen/cuda/nvrtc_stub/ATenNVRTC.h b/aten/src/ATen/cuda/nvrtc_stub/ATenNVRTC.h
+index 9cd678dfb4cc7..4630465115c7c 100644
+--- a/aten/src/ATen/cuda/nvrtc_stub/ATenNVRTC.h
++++ b/aten/src/ATen/cuda/nvrtc_stub/ATenNVRTC.h
+@@ -67,6 +67,14 @@ namespace at { namespace cuda {
+ //
+ // HIP doesn't have
+ // cuGetErrorString (maps to non-functional hipGetErrorString___)
++//
++// HIP from ROCm 3.5 on renamed hipOccupancyMaxActiveBlocksPerMultiprocessor
++// to hipModuleOccupancyMaxActiveBlocksPerMultiprocessor.
++#if HIP_VERSION < 305
++#define HIPOCCUPANCYMAXACTIVEBLOCKSPERMULTIPROCESSOR hipOccupancyMaxActiveBlocksPerMultiprocessor
++#else
++#define HIPOCCUPANCYMAXACTIVEBLOCKSPERMULTIPROCESSOR cuOccupancyMaxActiveBlocksPerMultiprocessor
++#endif
+
+ #define AT_FORALL_NVRTC(_) \
+ _(nvrtcVersion) \
+@@ -76,7 +84,7 @@ namespace at { namespace cuda {
+ _(nvrtcGetPTX) \
+ _(cuModuleLoadData) \
+ _(cuModuleGetFunction) \
+- _(cuOccupancyMaxActiveBlocksPerMultiprocessor) \
++ _(HIPOCCUPANCYMAXACTIVEBLOCKSPERMULTIPROCESSOR)\
+ _(nvrtcGetErrorString) \
+ _(nvrtcGetProgramLogSize) \
+ _(nvrtcGetProgramLog) \
+diff --git a/aten/src/ATen/native/cuda/SoftMax.cu b/aten/src/ATen/native/cuda/SoftMax.cu
+index da1995123ecfc..f935eb4ef3d0e 100644
+--- a/aten/src/ATen/native/cuda/SoftMax.cu
++++ b/aten/src/ATen/native/cuda/SoftMax.cu
+@@ -127,8 +127,8 @@ void SpatialSoftMax_getLaunchSizes(
+ uint32_t block_threads = block.x * block.y;
+ smem_size = block.x == 1 ? 0 : block_threads * sizeof(accscalar_t);
+ int max_active_blocks;
+-#ifdef __HIP_PLATFORM_HCC__
+- // XXX HIP function signature is not compatible yet.
++#if defined(__HIP_PLATFORM_HCC__) && HIP_VERSION < 305
++ // HIP function signature is not compatible yet.
+ uint32_t max_blocks;
+ cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks,
+ k, block_threads, smem_size);
+diff --git a/torch/csrc/jit/codegen/fuser/cuda/fused_kernel.cpp b/torch/csrc/jit/codegen/fuser/cuda/fused_kernel.cpp
+index 5586e49919727..27315ee475277 100644
+--- a/torch/csrc/jit/codegen/fuser/cuda/fused_kernel.cpp
++++ b/torch/csrc/jit/codegen/fuser/cuda/fused_kernel.cpp
+@@ -140,10 +140,10 @@ FusedKernelCUDA::FusedKernelCUDA(
+ nvrtc().cuModuleGetFunction(&function_, module_, name_.c_str()));
+
+ // Computes max blocks
+-#ifdef __HIP_PLATFORM_HCC__
+- // XXX HIP function signature is not compatible yet
++#if defined(__HIP_PLATFORM_HCC__) && HIP_VERSION < 305
++ // HIP function signature is not compatible yet
+ uint32_t max_blocks;
+- AT_CUDA_DRIVER_CHECK(nvrtc().cuOccupancyMaxActiveBlocksPerMultiprocessor(
++ AT_CUDA_DRIVER_CHECK(nvrtc().hipOccupancyMaxActiveBlocksPerMultiprocessor(
+ &max_blocks, function_, 128, 0));
+ maxBlocks_ = max_blocks;
+ #else
+diff --git a/torch/utils/hipify/cuda_to_hip_mappings.py b/torch/utils/hipify/cuda_to_hip_mappings.py
+index 7e21363cbe6af..26f269d92ae38 100644
+--- a/torch/utils/hipify/cuda_to_hip_mappings.py
++++ b/torch/utils/hipify/cuda_to_hip_mappings.py
+@@ -2890,7 +2890,7 @@
+ (
+ "cuOccupancyMaxActiveBlocksPerMultiprocessor",
+ (
+- "hipOccupancyMaxActiveBlocksPerMultiprocessor",
++ "hipModuleOccupancyMaxActiveBlocksPerMultiprocessor",
+ CONV_OCCUPANCY,
+ API_DRIVER,
+ ),
+@@ -2898,7 +2898,7 @@
+ (
+ "cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags",
+ (
+- "hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags",
++ "hipModuleOccupancyMaxActiveBlocksPerMultiprocessorWithFlags",
+ CONV_OCCUPANCY,
+ API_DRIVER,
+ HIP_UNSUPPORTED,
+@@ -2906,12 +2906,12 @@
+ ),
+ (
+ "cuOccupancyMaxPotentialBlockSize",
+- ("hipOccupancyMaxPotentialBlockSize", CONV_OCCUPANCY, API_DRIVER),
++ ("hipModuleOccupancyMaxPotentialBlockSize", CONV_OCCUPANCY, API_DRIVER),
+ ),
+ (
+ "cuOccupancyMaxPotentialBlockSizeWithFlags",
+ (
+- "hipOccupancyMaxPotentialBlockSizeWithFlags",
++ "hipModuleOccupancyMaxPotentialBlockSizeWithFlags",
+ CONV_OCCUPANCY,
+ API_DRIVER,
+ HIP_UNSUPPORTED,
diff --git a/packages/py-torch/sleef.patch b/packages/py-torch/sleef.patch
new file mode 100644
index 0000000000000000000000000000000000000000..67f0234162d1a1af29aa5c538b0e585c3261a81e
--- /dev/null
+++ b/packages/py-torch/sleef.patch
@@ -0,0 +1,12 @@
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 8025a7de3c..2e5cdbb5c9 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1232,6 +1232,7 @@ if (BUILD_TEST)
+ add_executable(${test_name} "${test_src}")
+ target_link_libraries(${test_name} ${Caffe2_MAIN_LIBS} gtest_main)
+ target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
++ target_include_directories(${test_name} PRIVATE $<BUILD_INTERFACE:${CMAKE_BINARY_DIR}/include>)
+ target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE})
+ add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+ if (INSTALL_TEST)
diff --git a/packages/py-torch/xnnpack.patch b/packages/py-torch/xnnpack.patch
new file mode 100644
index 0000000000000000000000000000000000000000..154033081e7ff91867e9a043a93c46b888bfe8cb
--- /dev/null
+++ b/packages/py-torch/xnnpack.patch
@@ -0,0 +1,47 @@
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 8025a7de3c..0da37079d6 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -46,12 +46,19 @@ if (INTERN_BUILD_ATEN_OPS)
+ list(APPEND Caffe2_DEPENDENCY_INCLUDE ${ATen_THIRD_PARTY_INCLUDE})
+ endif()
+
++# {Q/X,etc} NPACK support is enabled by default, if none of these options
++# are selected, turn this flag ON to incidate the support is disabled
++set(NNPACK_AND_FAMILY_DISABLED OFF)
++if(NOT (USE_NNPACK OR USE_QNNPACK OR USE_PYTORCH_QNNPACK OR USE_XNNPACK))
++ set(NNPACK_AND_FAMILY_DISABLED ON)
++endif()
++
+ # ---[ Caffe2 build
+ # Note: the folders that are being commented out have not been properly
+ # addressed yet.
+
+ # For pthreadpool_new_if_impl. TODO: Remove when threadpools are unitied.
+-if (NOT MSVC)
++if (NOT MSVC AND NOT NNPACK_AND_FAMILY_DISABLED)
+ IF(NOT TARGET fxdiv)
+ SET(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+ SET(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+@@ -710,7 +717,7 @@ ELSEIF(USE_CUDA)
+ ENDIF()
+
+
+-if (NOT MSVC)
++if (NOT MSVC AND NOT NNPACK_AND_FAMILY_DISABLED)
+ TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+ endif()
+
+diff --git a/caffe2/utils/CMakeLists.txt b/caffe2/utils/CMakeLists.txt
+index 27aabb1315..3c7845c67d 100644
+--- a/caffe2/utils/CMakeLists.txt
++++ b/caffe2/utils/CMakeLists.txt
+@@ -36,7 +36,7 @@ list(APPEND Caffe2_CPU_SRCS
+ # ---[ threadpool/pthreadpool* is a local modification of the NNPACK
+ # pthreadpool with a very similar interface. Neither NNPACK, nor this
+ # thread pool supports Windows.
+-if (NOT MSVC)
++if (NOT MSVC AND NOT NNPACK_AND_FAMILY_DISABLED)
+ add_definitions(-DUSE_INTERNAL_THREADPOOL_IMPL)
+ set(Caffe2_CPU_SRCS ${Caffe2_CPU_SRCS}
+ utils/threadpool/pthreadpool.cc