gnu: Add python-pytorch.

* gnu/packages/machine-learning.scm (python-pytorch): New variable.
* gnu/packages/patches/python-pytorch-runpath.patch,
gnu/packages/patches/python-pytorch-system-libraries.patch: New files.
* gnu/local.mk (dist_patch_DATA): Add them.

gnu/local.mk

@@ -1649,6 +1649,8 @@ dist_patch_DATA =						\
   %D%/packages/patches/python-pydot-regression-test.patch	\
   %D%/packages/patches/python2-pygobject-2-deprecation.patch	\
   %D%/packages/patches/python-pygpgme-fix-pinentry-tests.patch	\
+  %D%/packages/patches/python-pytorch-runpath.patch		\
+  %D%/packages/patches/python-pytorch-system-libraries.patch	\
   %D%/packages/patches/python-robotframework-source-date-epoch.patch \
   %D%/packages/patches/python-seaborn-kde-test.patch		\
   %D%/packages/patches/python2-subprocess32-disable-input-test.patch	\

gnu/packages/machine-learning.scm

@@ -65,10 +65,12 @@
   #:use-module (gnu packages graphviz)
   #:use-module (gnu packages gstreamer)
   #:use-module (gnu packages image)
+  #:use-module (gnu packages libffi)
   #:use-module (gnu packages linux)
   #:use-module (gnu packages llvm)
   #:use-module (gnu packages maths)
   #:use-module (gnu packages mpi)
+  #:use-module (gnu packages ninja)
   #:use-module (gnu packages ocaml)
   #:use-module (gnu packages onc-rpc)
   #:use-module (gnu packages parallel)
@@ -2546,7 +2548,7 @@ general non-linear dimension reduction.")
     (package
       (name "xnnpack")
       (version (git-version version revision commit))
-      (home-page "https://github.com/google/XNNPACK")
+      (home-page "https://github.com/google/XNNPACK") ;fork of QNNPACK
       (source (origin
                 (method git-fetch)
                 (uri (git-reference (url home-page) (commit commit)))
@@ -2586,3 +2588,142 @@ high-level machine learning frameworks, such as TensorFlow Lite,
 TensorFlow.js, PyTorch, and MediaPipe.")
       (license license:bsd-3))))
 
+(define-public python-pytorch
+  (package
+    (name "python-pytorch")
+    (version "1.9.0")
+    (source (origin
+              (method git-fetch)
+              (uri (git-reference
+                    (url "https://github.com/pytorch/pytorch")
+                    (commit (string-append "v" version))
+                    (recursive? #t)))
+              (file-name (git-file-name name version))
+              (sha256
+               (base32
+                "0cznsh68hwk5761gv7iijb4g6jgjpvs3bbixwpzzmkbkbn2q96c1"))
+              (patches (search-patches "python-pytorch-system-libraries.patch"
+                                       "python-pytorch-runpath.patch"))
+              (modules '((guix build utils)))
+              (snippet
+               '(begin
+                  ;; XXX: Let's be clear: this package is a bundling fest.  We
+                  ;; delete as much as we can, but there's still a lot left.
+                  (for-each (lambda (directory)
+                              (delete-file-recursively
+                               (string-append "third_party/" directory)))
+                            '("benchmark" "cpuinfo" "eigen"
+
+                              ;; FIXME: QNNPACK (of which XNNPACK is a fork)
+                              ;; needs these.
+                              ;; "FP16" "FXdiv" "gemmlowp" "psimd"
+
+                              "gloo" "googletest" "ios-cmake"
+                              "onnx" "protobuf" "pthreadpool"
+                              "pybind11" "python-enum" "python-peachpy"
+                              "python-six" "tbb" "XNNPACK" "zstd"))
+
+                  ;; Adjust references to the onnx-optimizer headers.
+                  (substitute* "caffe2/onnx/backend.cc"
+                    (("onnx/optimizer/")
+                     "onnxoptimizer/"))))))
+    (build-system python-build-system)
+    (arguments
+     '(#:phases (modify-phases %standard-phases
+                  (add-before 'build 'use-system-libraries
+                    (lambda* (#:key outputs #:allow-other-keys)
+                      ;; Tell 'setup.py' to let 'CMakeLists.txt' know that we
+                      ;; want to use "system libraries" instead of the bundled
+                      ;; ones.
+                      (setenv "USE_SYSTEM_LIBS" "1")
+
+                      ;; XXX: Disable that for simplicity for now.
+                      (setenv "USE_FBGEMM" "0")
+
+                      ;; Set 'OUT/lib' rather than '$ORIGIN' as the RUNPATH on
+                      ;; binaries that go to 'lib/python3.8/torch/bin' & co.
+                      #;(substitute* "cmake/Dependencies.cmake"
+                        (("^set\\(CMAKE_INSTALL_RPATH .*")
+                         (string-append "set(CMAKE_INSTALL_RPATH \""
+                                        (assoc-ref outputs "out")
+                                        "/lib\")\n")))))
+                  (add-before 'build 'make-things-writable
+                    (lambda _
+                      ;; The 'build_caffe2' function in
+                      ;; 'tools/build_pytorch_libs.py', called from the
+                      ;; top-level 'setup.py', needs write access to this
+                      ;; directory.
+                      (for-each make-file-writable
+                                (find-files "caffe2/proto" "."
+                                            #:directories? #t))))
+                  (replace 'check
+                    (lambda* (#:key inputs outputs tests? #:allow-other-keys)
+                      ;; Run the test suite following the instructions in
+                      ;; 'CONTRIBUTING.md'.  XXX: Unfortunately this doesn't
+                      ;; work, unless you set PYTHONPATH presumably.
+                      (when tests?
+                        (let ((python-site (site-packages inputs outputs)))
+                          (setenv "PYTHONPATH"
+                                  (string-append python-site ":"
+                                                 (getenv "PYTHONPATH")))
+                          (invoke "python" "test/run_test.py")))))
+                  (add-after 'install 'remove-test-executables
+                    (lambda* (#:key inputs outputs #:allow-other-keys)
+                      ;; Remove test executables, but keep other executables
+                      ;; such as 'torch_shm_manager' and and .so files such as
+                      ;; 'libtorch_global_deps.so'.
+                      (let ((python-site (site-packages inputs outputs)))
+                        (for-each delete-file
+                                  (find-files python-site
+                                              "(^test_cpp_rpc|_test)$"))))))
+
+       ;; XXX: Tests attempt to download data such as
+       ;; <https://raw.githubusercontent.com/pytorch/test-infra/master/stats/slow-tests.json>.
+       #:tests? #f))
+    (native-inputs
+     `(("cmake" ,cmake)
+       ("ninja" ,ninja)))
+    (inputs
+     `(("eigen" ,eigen)
+       ;; ("fmt" ,fmt)
+       ("fp16" ,fp16)
+       ("gemmlowp" ,gemmlowp)
+       ("googletest" ,googletest)
+       ("googlebenchmark" ,googlebenchmark)
+       ("gloo" ,gloo)
+       ("openblas" ,openblas)
+       ("openmpi" ,openmpi)
+       ("pthreadpool" ,pthreadpool)
+       ("protobuf" ,protobuf)
+       ("pybind11" ,pybind11)
+       ("sleef" ,sleef)
+       ("xnnpack" ,xnnpack)
+       ("zstd" ,zstd)))
+    (propagated-inputs
+     `(("python-astunparse" ,python-astunparse)
+       ("python-numpy" ,python-numpy)
+       ("python-pyyaml" ,python-pyyaml)
+       ("python-cffi" ,python-cffi)
+       ("python-peachpy" ,python-peachpy)
+       ("python-typing-extensions" ,python-typing-extensions)
+       ("python-future" ,python-future)
+       ("python-six" ,python-six)
+       ("python-requests" ,python-requests)
+       ("onnx" ,onnx)                       ;propagated for its Python modules
+       ("onnx-optimizer" ,onnx-optimizer)
+       ("cpuinfo" ,cpuinfo)))
+    (home-page "https://pytorch.org/")
+    (synopsis "Python library for tensor computation and deep neural networks")
+    (description
+     "PyTorch is a Python package that provides two high-level features:
+
+@itemize
+@item tensor computation (like NumPy) with strong GPU acceleration;
+@item deep neural networks (DNNs) built on a tape-based autograd system.
+@end itemize
+
+You can reuse Python packages such as NumPy, SciPy, and Cython to extend
+PyTorch when needed.
+
+Note: currently this package does not provide GPU support.")
+    (license license:bsd-3)))

gnu/packages/patches/python-pytorch-runpath.patch

@@ -0,0 +1,25 @@
+Libraries (such as 'libtorch_cpu.so') and executables (such as 'torch_shm_manager')
+get installed, quite surprisingly, to 'lib/python3.8/site-packages/{bin,lib}'.
+Make sure RUNPATH matches that.
+
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -4,7 +4,7 @@ if(APPLE)
+   set(CMAKE_MACOSX_RPATH ON)
+   set(_rpath_portable_origin "@loader_path")
+ else()
+-  set(_rpath_portable_origin $ORIGIN)
++  set(_rpath_portable_origin $ORIGIN/../lib)
+ endif(APPLE)
+ # Use separate rpaths during build and install phases
+ set(CMAKE_SKIP_BUILD_RPATH  FALSE)
+
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1797,5 +1797,5 @@ if(BUILD_PYTHON)
+   if(${BUILDING_WITH_TORCH_LIBS})
+     # site-packages/caffe2/python/caffe2_pybind11_state
+     # site-packages/torch/lib
+-    set(caffe2_pybind11_rpath "${_rpath_portable_origin}/../../torch/lib")
++    set(caffe2_pybind11_rpath $ORIGIN/../../torch/lib)
+   endif(${BUILDING_WITH_TORCH_LIBS})

gnu/packages/patches/python-pytorch-system-libraries.patch

@@ -0,0 +1,131 @@
+Use our own googletest rather than the bundled one.
+Get NNPACK to use our own PeachPy rather than the bundled one.
+
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 5d57b9ca78..620cca4e60 100644
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -644,11 +644,6 @@ if(BUILD_TEST OR BUILD_MOBILE_BENCHMARK OR BUILD_MOBILE_TEST)
+   # this shouldn't be necessary anymore.
+   get_property(INC_DIR_temp DIRECTORY PROPERTY INCLUDE_DIRECTORIES)
+   set_property(DIRECTORY PROPERTY INCLUDE_DIRECTORIES "")
+-  add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/../third_party/googletest)
+-  set_property(DIRECTORY PROPERTY INCLUDE_DIRECTORIES ${INC_DIR_temp})
+-
+-  include_directories(BEFORE SYSTEM ${CMAKE_CURRENT_LIST_DIR}/../third_party/googletest/googletest/include)
+-  include_directories(BEFORE SYSTEM ${CMAKE_CURRENT_LIST_DIR}/../third_party/googletest/googlemock/include)
+ 
+   # We will not need to test benchmark lib itself.
+   set(BENCHMARK_ENABLE_TESTING OFF CACHE BOOL "Disable benchmark testing as we don't need it.")
+@@ -1485,7 +1480,7 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+     endif()
+     set_property(TARGET onnx_proto PROPERTY IMPORTED_LOCATION ${ONNX_PROTO_LIBRARY})
+     message("-- Found onnx: ${ONNX_LIBRARY} ${ONNX_PROTO_LIBRARY}")
+-    list(APPEND Caffe2_DEPENDENCY_LIBS onnx_proto onnx)
++    list(APPEND Caffe2_DEPENDENCY_LIBS onnx_proto onnx onnx_optimizer)
+   endif()
+   include_directories(${FOXI_INCLUDE_DIRS})
+   list(APPEND Caffe2_DEPENDENCY_LIBS foxi_loader)
+
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 50ebb224ce..5953d9ddf7 100644
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1632,7 +1632,7 @@ if(BUILD_TEST)
+         if(NOT MSVC)
+           add_executable(${test_name}_${CPU_CAPABILITY} "${test_src}" ../aten/src/ATen/native/quantized/affine_quantizer_base.cpp)
+           # TODO: Get rid of c10 dependency (which is only needed for the implementation of AT_ERROR)
+-          target_link_libraries(${test_name}_${CPU_CAPABILITY} c10 sleef gtest_main)
++          target_link_libraries(${test_name}_${CPU_CAPABILITY} c10 sleef gtest_main gtest)
+           if(USE_FBGEMM)
+             target_link_libraries(${test_name}_${CPU_CAPABILITY} fbgemm)
+           endif()
+@@ -1655,7 +1655,7 @@ if(BUILD_TEST)
+   foreach(test_src ${Caffe2_CPU_TEST_SRCS})
+     get_filename_component(test_name ${test_src} NAME_WE)
+     add_executable(${test_name} "${test_src}")
+-    target_link_libraries(${test_name} torch_library gtest_main)
++    target_link_libraries(${test_name} torch_library gtest_main gtest)
+     target_include_directories(${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})
+@@ -1673,7 +1673,7 @@ if(BUILD_TEST)
+     foreach(test_src ${Caffe2_GPU_TEST_SRCS})
+       get_filename_component(test_name ${test_src} NAME_WE)
+       cuda_add_executable(${test_name} "${test_src}")
+-      target_link_libraries(${test_name} torch_library gtest_main)
++      target_link_libraries(${test_name} torch_library gtest_main gtest)
+       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
+       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE})
+       add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+@@ -1691,7 +1691,7 @@ if(BUILD_TEST)
+     foreach(test_src ${Caffe2_VULKAN_TEST_SRCS})
+       get_filename_component(test_name ${test_src} NAME_WE)
+       add_executable(${test_name} "${test_src}")
+-      target_link_libraries(${test_name} torch_library gtest_main)
++      target_link_libraries(${test_name} torch_library gtest_main gtest)
+       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
+       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE})
+       add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+@@ -1709,7 +1709,7 @@ if(BUILD_TEST)
+     foreach(test_src ${Caffe2_HIP_TEST_SRCS})
+       get_filename_component(test_name ${test_src} NAME_WE)
+       add_executable(${test_name} "${test_src}")
+-      target_link_libraries(${test_name} torch_library gtest_main)
++      target_link_libraries(${test_name} torch_library gtest_main gtest)
+       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
+       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE} ${Caffe2_HIP_INCLUDE})
+       target_compile_options(${test_name} PRIVATE ${HIP_CXX_FLAGS})
+
+diff --git a/torch/lib/c10d/test/CMakeLists.txt b/torch/lib/c10d/test/CMakeLists.txt
+index b74d4b65f7..fc7c207505 100644
+--- a/torch/lib/c10d/test/CMakeLists.txt
++++ b/torch/lib/c10d/test/CMakeLists.txt
+@@ -16,25 +16,25 @@ function(c10d_add_test test_src)
+   add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+ endfunction()
+ 
+-c10d_add_test(FileStoreTest.cpp c10d gtest_main)
+-c10d_add_test(TCPStoreTest.cpp c10d gtest_main)
++c10d_add_test(FileStoreTest.cpp c10d gtest_main gtest)
++c10d_add_test(TCPStoreTest.cpp c10d gtest_main gtest)
+ if(NOT WIN32)
+-  c10d_add_test(HashStoreTest.cpp c10d gtest_main)
++  c10d_add_test(HashStoreTest.cpp c10d gtest_main gtest)
+ endif()
+ 
+ if(USE_CUDA)
+   if(USE_C10D_GLOO)
+-    c10d_add_test(ProcessGroupGlooTest.cpp c10d c10d_cuda_test gtest_main)
+-    c10d_add_test(ProcessGroupGlooAsyncTest.cpp c10d c10d_cuda_test gtest_main)
++    c10d_add_test(ProcessGroupGlooTest.cpp c10d c10d_cuda_test gtest_main gtest)
++    c10d_add_test(ProcessGroupGlooAsyncTest.cpp c10d c10d_cuda_test gtest_main gtest)
+   endif()
+   if(USE_C10D_NCCL)
+-    c10d_add_test(ProcessGroupNCCLTest.cpp c10d c10d_cuda_test gtest_main)
++    c10d_add_test(ProcessGroupNCCLTest.cpp c10d c10d_cuda_test gtest_main gtest)
+     c10d_add_test(ProcessGroupNCCLErrorsTest.cpp c10d c10d_cuda_test
+-        gtest_main)
++        gtest_main gtest)
+   endif()
+ else()
+   if(USE_C10D_GLOO)
+-    c10d_add_test(ProcessGroupGlooTest.cpp c10d gtest_main)
++    c10d_add_test(ProcessGroupGlooTest.cpp c10d gtest_main gtest)
+   endif()
+ endif()
+ 
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 5ecd2df..24feae3 100644
+--- a/third_party/NNPACK/CMakeLists.txt
++++ b/third_party/NNPACK/CMakeLists.txt
+@@ -427,8 +427,7 @@ IF(NNPACK_BACKEND STREQUAL "x86-64")
+       FILE(MAKE_DIRECTORY ${obj_dir})
+       ADD_CUSTOM_COMMAND(
+         OUTPUT ${obj}
+-        COMMAND "PYTHONPATH=${PEACHPY_PYTHONPATH}"
+-          ${PYTHON_EXECUTABLE} -m peachpy.x86_64
++        COMMAND ${PYTHON_EXECUTABLE} -m peachpy.x86_64
+             -mabi=sysv -g4 -mimage-format=${PEACHPY_IMAGE_FORMAT}
+             "-I${PROJECT_SOURCE_DIR}/src" "-I${PROJECT_SOURCE_DIR}/src/x86_64-fma" "-I${FP16_SOURCE_DIR}/include"
+             -o ${obj} "${PROJECT_SOURCE_DIR}/${src}"