[libtorch] update to 2.1.2 (#36246)

* [libtorch] update to 2.1.2 * adjust features and deps * remove dep * fix onnx case * use system fxdiv * update xnnpack * fxdiv fix in libtorch * add newline * xnnpack does not install configs * more patching * fix patch * more fixes * use full lib for sleef. * fix glog symbol on osx * try fixing msvc ice * ok renaming did nothing * v db * remove comments * v db * cuda is only for x64 * v db * fix usage since cuda is still external to vcpkg * v db --------- Co-authored-by: Jim wang (BEYONDSOFT CONSULTING INC) <v-wangjim@microsoft.com>
2024-11-27 19:59:06 +08:00 · 2024-01-30 23:15:33 +01:00 · 2024-01-30 23:15:33 +01:00 · 20584b9d44
commit 20584b9d44
parent 7233110263
32 changed files with 1376 additions and 1173 deletions
--- a/ports/libtorch/clang-cl.patch
+++ b/ports/libtorch/clang-cl.patch
@ -0,0 +1,47 @@
+diff --git a/c10/util/Exception.h b/c10/util/Exception.h
+index 2dd1001..b3f3d60 100644
+--- a/c10/util/Exception.h
+++ b/c10/util/Exception.h
+@@ -233,12 +233,16 @@ struct C10_API WarnAlways {
+ // IndexError when they cross to Python.
+ class C10_API IndexError : public Error {
+   using Error::Error;
+public:
+  inline IndexError(SourceLocation source_location, std::string msg) : Error(source_location,msg) {};
+ };
+ 
+ // Used in ATen for invalid values.  These turn into
+ // ValueError when they cross to Python.
+ class C10_API ValueError : public Error {
+   using Error::Error;
+public:
+  inline ValueError(SourceLocation source_location, std::string msg) : Error(source_location,msg) {};
+ };
+ 
+ // Used in ATen for invalid types.  These turn into
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index cf349c2..f7a595e 100644
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -830,7 +830,7 @@ if(BUILD_CAFFE2 AND NOT MSVC)
+   target_compile_options(torch_cpu PRIVATE "-Wno-sign-compare")
+ endif()
+ 
+-if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang" AND NOT USE_VULKAN AND NOT USE_IOS AND NOT USE_PYTORCH_METAL AND NOT USE_COREML_DELEGATE)
+if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang" AND NOT USE_VULKAN AND NOT USE_IOS AND NOT USE_PYTORCH_METAL AND NOT USE_COREML_DELEGATE AND NOT MSVC)
+   target_compile_options_if_supported(torch_cpu "-Wmissing-prototypes")
+   target_compile_options_if_supported(torch_cpu "-Werror=missing-prototypes")
+   get_target_property(TORCH_CPU_SOURCES torch_cpu SOURCES)
+diff --git a/functorch/csrc/dim/arena.h b/functorch/csrc/dim/arena.h
+index 3251321..cb09b7c 100644
+--- a/functorch/csrc/dim/arena.h
+++ b/functorch/csrc/dim/arena.h
+@@ -8,7 +8,7 @@
+ #include <ATen/ATen.h>
+ #include "minpybind.h"
+ 
+-#ifdef _WIN32
+#if defined(_WIN32) && !defined(__clang__)
+ #include <intrin.h>
+ // https://stackoverflow.com/questions/355967/how-to-use-msvc-intrinsics-to-get-the-equivalent-of-this-gcc-code
+ inline unsigned int __builtin_clz(unsigned int x) {
--- a/ports/libtorch/cmake-fixes.patch
+++ b/ports/libtorch/cmake-fixes.patch
@ -0,0 +1,235 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 3a48eaf4e2..7b8bc7cb6e 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -1049,8 +1049,7 @@ if(USE_MIMALLOC)
+   set(MI_BUILD_OBJECT OFF)
+   set(MI_BUILD_TESTS OFF)
+   add_definitions(-DUSE_MIMALLOC)
+-  add_subdirectory(third_party/mimalloc)
+-  include_directories(third_party/mimalloc/include)
+  find_package(mimalloc REQUIRED)
+ endif()
+ 
+ # ---[ Main build
+diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+index fd6b7ff551..d185233529 100644
+--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+@@ -368,7 +368,7 @@ elseif(NOT TARGET pthreadpool AND USE_SYSTEM_PTHREADPOOL)
+   if(NOT PTHREADPOOL_LIBRARY)
+     message(FATAL_ERROR "Cannot find pthreadpool")
+   endif()
+-  message("-- Found pthreadpool: ${PTHREADPOOL_LIBRARY}")
+  message("-- Found pthreadpool cpu qnnpack: ${PTHREADPOOL_LIBRARY}")
+   set_target_properties(pthreadpool PROPERTIES
+     IMPORTED_LOCATION "${PTHREADPOOL_LIBRARY}")
+   add_library(pthreadpool_interface INTERFACE)
+diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
+index feebad7cbb..ad407e18cb 100644
+--- a/c10/CMakeLists.txt
+++ b/c10/CMakeLists.txt
+@@ -109,8 +109,7 @@ else()
+ endif()
+ 
+ if(USE_MIMALLOC)
+-  target_link_libraries(c10 PRIVATE "mimalloc-static")
+-  add_dependencies(c10 mimalloc-static)
+  target_link_libraries(c10 PRIVATE "mimalloc")
+ endif()
+ 
+ if(ANDROID)
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index c3abce52e4..cd27c728ff 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -285,7 +285,7 @@ endif()
+ set(AT_FFTW_ENABLED 0)
+ set(USE_FFTW OFF)
+ if(USE_FFTW OR NOT MKL_FOUND)
+-  find_library(LIBFFTW3 fftw3)
+  find_library(LIBFFTW3 NAMES fftw3 REQUIRED)
+   if(LIBFFTW3)
+     find_path(FFTW3_INCLUDE_DIR NAMES fftw3.h ONLY_CMAKE_FIND_ROOT_PATH)
+     if(FFTW3_INCLUDE_DIR)
+@@ -415,8 +415,8 @@ if(INTERN_BUILD_MOBILE OR NOT DISABLE_NNPACK_AND_FAMILY)
+ 
+   if(NOT TARGET pthreadpool)
+     if(USE_SYSTEM_PTHREADPOOL)
+-      add_library(pthreadpool SHARED IMPORTED)
+-      find_library(PTHREADPOOL_LIBRARY pthreadpool)
+      add_library(pthreadpool UNKNOWN IMPORTED)
+      find_library(PTHREADPOOL_LIBRARY pthreadpool PATH_SUFFIXES lib)
+       set_property(TARGET pthreadpool PROPERTY IMPORTED_LOCATION "${PTHREADPOOL_LIBRARY}")
+       if(NOT PTHREADPOOL_LIBRARY)
+         message(FATAL_ERROR "Cannot find pthreadpool")
+@@ -450,13 +450,15 @@ endif()
+ 
+ # ---[ Caffe2 uses cpuinfo library in the thread pool
+ if(NOT TARGET cpuinfo AND USE_SYSTEM_CPUINFO)
+-  add_library(cpuinfo SHARED IMPORTED)
+-  find_library(CPUINFO_LIBRARY cpuinfo)
+  add_library(cpuinfo UNKNOWN IMPORTED)
+  find_library(CPUINFO_LIBRARY NAMES cpuinfo REQUIRED)
+  find_library(CLOG_LIBRARY NAMES clog REQUIRED)
+   if(NOT CPUINFO_LIBRARY)
+     message(FATAL_ERROR "Cannot find cpuinfo")
+   endif()
+   message("Found cpuinfo: ${CPUINFO_LIBRARY}")
+   set_target_properties(cpuinfo PROPERTIES IMPORTED_LOCATION "${CPUINFO_LIBRARY}")
+  target_link_libraries(cpuinfo INTERFACE "${CLOG_LIBRARY}")
+ elseif(NOT TARGET cpuinfo)
+   if(NOT DEFINED CPUINFO_SOURCE_DIR)
+     set(CPUINFO_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/../third_party/cpuinfo" CACHE STRING "cpuinfo source directory")
+@@ -659,8 +661,8 @@ if(USE_XNNPACK AND NOT USE_SYSTEM_XNNPACK)
+   include_directories(SYSTEM ${XNNPACK_INCLUDE_DIR})
+   list(APPEND Caffe2_DEPENDENCY_LIBS XNNPACK)
+ elseif(NOT TARGET XNNPACK AND USE_SYSTEM_XNNPACK)
+-  add_library(XNNPACK SHARED IMPORTED)
+-  find_library(XNNPACK_LIBRARY XNNPACK)
+  add_library(XNNPACK UNKNOWN IMPORTED)
+  find_library(XNNPACK_LIBRARY NAMES XNNPACK REQUIRED)
+   set_property(TARGET XNNPACK PROPERTY IMPORTED_LOCATION "${XNNPACK_LIBRARY}")
+   if(NOT XNNPACK_LIBRARY)
+     message(FATAL_ERROR "Cannot find XNNPACK")
+@@ -858,7 +860,7 @@ endif()
+ 
+ # ---[ LMDB
+ if(USE_LMDB)
+-  find_package(LMDB)
+  find_package(LMDB REQUIRED)
+   if(LMDB_FOUND)
+     include_directories(SYSTEM ${LMDB_INCLUDE_DIR})
+     list(APPEND Caffe2_DEPENDENCY_LIBS ${LMDB_LIBRARIES})
+@@ -1002,8 +1004,9 @@ if(NOT TARGET fp16 AND NOT USE_SYSTEM_FP16)
+     "${FP16_SOURCE_DIR}"
+     "${CONFU_DEPENDENCIES_BINARY_DIR}/FP16")
+ elseif(NOT TARGET fp16 AND USE_SYSTEM_FP16)
+-  add_library(fp16 STATIC "/usr/include/fp16.h")
+-  set_target_properties(fp16 PROPERTIES LINKER_LANGUAGE C)
+  find_path(FP16_INCLUDE_DIR NAMES fp16.h PATH_SUFFIXES include)
+  add_library(fp16 INTERFACE)
+  target_include_directories(fp16 INTERFACE "${FP16_INCLUDE_DIR}")
+ endif()
+ list(APPEND Caffe2_DEPENDENCY_LIBS fp16)
+ 
+@@ -1244,8 +1247,8 @@ endif(USE_LLVM)
+ 
+ # ---[ cuDNN
+ if(USE_CUDNN)
+-  set(CUDNN_FRONTEND_INCLUDE_DIR ${CMAKE_CURRENT_LIST_DIR}/../third_party/cudnn_frontend/include)
+-  target_include_directories(torch::cudnn INTERFACE ${CUDNN_FRONTEND_INCLUDE_DIR})
+  find_package(CUDNN REQUIRED)
+  target_include_directories(torch::cudnn INTERFACE "${CUDNN_INCLUDE_DIR}")
+ endif()
+ 
+ # ---[ HIP
+@@ -1555,7 +1558,10 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+       set_target_properties(onnx_proto PROPERTIES CXX_STANDARD 17)
+     endif()
+   endif()
+-  add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/../third_party/foxi EXCLUDE_FROM_ALL)
+  
+  # setup foxi
+  find_library(FOXI_LOADER_LIBPATH NAMES foxi_loader REQUIRED)
+  list(APPEND Caffe2_DEPENDENCY_LIBS "${FOXI_LOADER_LIBPATH}")
+ 
+   add_definitions(-DONNX_NAMESPACE=${ONNX_NAMESPACE})
+   if(NOT USE_SYSTEM_ONNX)
+@@ -1573,14 +1579,14 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+       target_compile_options(onnx PRIVATE -Wno-deprecated-declarations)
+     endif()
+   else()
+-    add_library(onnx SHARED IMPORTED)
+-    find_library(ONNX_LIBRARY onnx)
+    add_library(onnx UNKNOWN IMPORTED)
+    find_library(ONNX_LIBRARY NAMES onnx REQUIRED)
+     if(NOT ONNX_LIBRARY)
+       message(FATAL_ERROR "Cannot find onnx")
+     endif()
+     set_property(TARGET onnx PROPERTY IMPORTED_LOCATION ${ONNX_LIBRARY})
+-    add_library(onnx_proto SHARED IMPORTED)
+-    find_library(ONNX_PROTO_LIBRARY onnx_proto)
+    add_library(onnx_proto UNKNOWN IMPORTED)
+    find_library(ONNX_PROTO_LIBRARY NAMES onnx_proto REQUIRED)
+     if(NOT ONNX_PROTO_LIBRARY)
+       message(FATAL_ERROR "Cannot find onnx")
+     endif()
+@@ -1589,7 +1595,7 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+     list(APPEND Caffe2_DEPENDENCY_LIBS onnx_proto onnx)
+   endif()
+   include_directories(${FOXI_INCLUDE_DIRS})
+-  list(APPEND Caffe2_DEPENDENCY_LIBS foxi_loader)
+  list(APPEND Caffe2_DEPENDENCY_LIBS "${FOXI_LOADER_LIBPATH}")
+   # Recover the build shared libs option.
+   set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS})
+ endif()
+@@ -1834,8 +1840,8 @@ endif()
+ #
+ set(TEMP_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+ set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build shared libs" FORCE)
+-add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+-
+#add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+find_package(fmt REQUIRED)
+ # Disable compiler feature checks for `fmt`.
+ #
+ # CMake compiles a little program to check compiler features. Some of our build
+@@ -1843,7 +1849,6 @@ add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+ # CMAKE_CXX_FLAGS in ways that break feature checks. Since we already know
+ # `fmt` is compatible with a superset of the compilers that PyTorch is, it
+ # shouldn't be too bad to just disable the checks.
+-set_target_properties(fmt-header-only PROPERTIES INTERFACE_COMPILE_FEATURES "")
+ 
+ list(APPEND Caffe2_DEPENDENCY_LIBS fmt::fmt-header-only)
+ set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
+diff --git a/cmake/Modules/FindLMDB.cmake b/cmake/Modules/FindLMDB.cmake
+index 2f0adb1b6d..8a817fd6f1 100644
+--- a/cmake/Modules/FindLMDB.cmake
+++ b/cmake/Modules/FindLMDB.cmake
+@@ -12,12 +12,8 @@
+ # Copyright 2013 Conrad Steenberg <conrad.steenberg@gmail.com>
+ # Aug 31, 2013
+ 
+-if(MSVC)
+-  find_package(LMDB NO_MODULE)
+-else()
+-  find_path(LMDB_INCLUDE_DIR NAMES  lmdb.h PATHS "$ENV{LMDB_DIR}/include")
+-  find_library(LMDB_LIBRARIES NAMES lmdb   PATHS "$ENV{LMDB_DIR}/lib" )
+-endif()
+find_path(LMDB_INCLUDE_DIR NAMES  lmdb.h PATHS "$ENV{LMDB_DIR}/include")
+find_library(LMDB_LIBRARIES NAMES lmdb   PATHS "$ENV{LMDB_DIR}/lib" )
+ 
+ include(FindPackageHandleStandardArgs)
+ find_package_handle_standard_args(LMDB DEFAULT_MSG LMDB_INCLUDE_DIR LMDB_LIBRARIES)
+diff --git a/cmake/public/cuda.cmake b/cmake/public/cuda.cmake
+index 32f3ba375b..411293794f 100644
+--- a/cmake/public/cuda.cmake
+++ b/cmake/public/cuda.cmake
+@@ -67,8 +67,8 @@ if(NOT CMAKE_CUDA_COMPILER_VERSION STREQUAL CUDAToolkit_VERSION OR
+                       "V${CUDAToolkit_VERSION} in '${CUDAToolkit_INCLUDE_DIR}'")
+ endif()
+ 
+-if(NOT TARGET CUDA::nvToolsExt)
+-  message(FATAL_ERROR "Failed to find nvToolsExt")
+if(NOT TARGET CUDA::nvToolsExt AND TARGET CUDA::nvtx3)
+  add_library(CUDA::nvToolsExt ALIAS CUDA::nvtx3)
+ endif()
+ 
+ message(STATUS "Caffe2: CUDA detected: " ${CUDA_VERSION})
+diff --git a/cmake/public/utils.cmake b/cmake/public/utils.cmake
+index 4d48c0f0f3..ebdd39a7fd 100644
+--- a/cmake/public/utils.cmake
+++ b/cmake/public/utils.cmake
+@@ -185,9 +185,9 @@ endfunction()
+ macro(caffe2_update_option variable value)
+   if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO)
+     get_property(__help_string CACHE ${variable} PROPERTY HELPSTRING)
+-    set(${variable} ${value} CACHE BOOL ${__help_string} FORCE)
+    set("${variable}" "${value}" CACHE BOOL "${__help_string}" FORCE)
+   else()
+-    set(${variable} ${value})
+    set("${variable}" "${value}")
+   endif()
+ endmacro()
+ 
--- a/ports/libtorch/cuda-adjustments.patch
+++ b/ports/libtorch/cuda-adjustments.patch
@ -0,0 +1,61 @@
+diff --git a/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake b/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
+index 9293df3..11bc4bf 100644
+--- a/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
+++ b/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
+@@ -126,11 +126,11 @@ list(APPEND CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS_${build_configuration}})
+ # Any -ccbin existing in CUDA_NVCC_FLAGS gets highest priority
+ list( FIND CUDA_NVCC_FLAGS "-ccbin" ccbin_found0 )
+ list( FIND CUDA_NVCC_FLAGS "--compiler-bindir" ccbin_found1 )
+-if( ccbin_found0 LESS 0 AND ccbin_found1 LESS 0 AND CUDA_HOST_COMPILER )
+if( ccbin_found0 LESS 0 AND ccbin_found1 LESS 0 AND CUDA_HOST_COMPILER AND 0)
+   if (CUDA_HOST_COMPILER STREQUAL "@_CUDA_MSVC_HOST_COMPILER@" AND DEFINED CCBIN)
+-    set(CCBIN -ccbin "${CCBIN}")
+    #set(CCBIN -ccbin "${CCBIN}")
+   else()
+-    set(CCBIN -ccbin "${CUDA_HOST_COMPILER}")
+    #set(CCBIN -ccbin "${CUDA_HOST_COMPILER}")
+   endif()
+ endif()
+ 
+diff --git a/cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake b/cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake
+index d917738..8f832fe 100644
+--- a/cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake
+++ b/cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake
+@@ -30,7 +30,7 @@ endif()
+ set(CUDA_KNOWN_GPU_ARCHITECTURES  "Kepler" "Maxwell")
+ 
+ # This list will be used for CUDA_ARCH_NAME = Common option (enabled by default)
+-set(CUDA_COMMON_GPU_ARCHITECTURES "3.5" "5.0")
+# set(CUDA_COMMON_GPU_ARCHITECTURES "3.5" "5.0")
+ 
+ # This list is used to filter CUDA archs when autodetecting
+ set(CUDA_ALL_GPU_ARCHITECTURES "3.5" "5.0")
+diff --git a/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake b/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
+index 420ee63..f029b66 100644
+--- a/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
+++ b/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
+@@ -1694,9 +1694,9 @@ function(CUDA_LINK_SEPARABLE_COMPILATION_OBJECTS output_file cuda_target options
+     if( ccbin_found0 LESS 0 AND ccbin_found1 LESS 0 AND CUDA_HOST_COMPILER )
+       # Match VERBATIM check below.
+       if(CUDA_HOST_COMPILER MATCHES "\\$\\(VCInstallDir\\)")
+-        list(APPEND nvcc_flags -ccbin "\"${CUDA_HOST_COMPILER}\"")
+        #list(APPEND nvcc_flags -ccbin "\"${CUDA_HOST_COMPILER}\"")
+       else()
+-        list(APPEND nvcc_flags -ccbin "${CUDA_HOST_COMPILER}")
+        #list(APPEND nvcc_flags -ccbin "${CUDA_HOST_COMPILER}")
+       endif()
+     endif()
+ 
+diff --git a/cmake/public/cuda.cmake b/cmake/public/cuda.cmake
+index 4112937..3e1af83 100644
+--- a/cmake/public/cuda.cmake
+++ b/cmake/public/cuda.cmake
+@@ -42,7 +42,7 @@ set(CUDAToolkit_ROOT "${CUDA_TOOLKIT_ROOT_DIR}")
+ # Must be done before CUDA language is enabled, see
+ # https://cmake.org/cmake/help/v3.15/variable/CMAKE_CUDA_HOST_COMPILER.html
+ if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
+-  set(CMAKE_CUDA_HOST_COMPILER "${CMAKE_C_COMPILER}")
+  #set(CMAKE_CUDA_HOST_COMPILER "${CMAKE_C_COMPILER}")
+ endif()
+ enable_language(CUDA)
+ if("X${CMAKE_CUDA_STANDARD}" STREQUAL "X" )
--- a/ports/libtorch/fix-api-export.patch
+++ b/ports/libtorch/fix-api-export.patch
@ -0,0 +1,39 @@
+diff --git a/aten/src/ATen/Context.h b/aten/src/ATen/Context.h
+index 285c6f7..1040acc 100644
+--- a/aten/src/ATen/Context.h
+++ b/aten/src/ATen/Context.h
+@@ -28,7 +28,7 @@ namespace at {
+ 
+ class Tensor;
+ 
+-enum class TORCH_API Float32MatmulPrecision { HIGHEST, HIGH, MEDIUM };
+enum class Float32MatmulPrecision { HIGHEST, HIGH, MEDIUM };
+ 
+ class TORCH_API Context {
+  public:
+diff --git a/aten/src/ATen/core/function_schema.h b/aten/src/ATen/core/function_schema.h
+index c48e9b0..cd8d4dd 100644
+--- a/aten/src/ATen/core/function_schema.h
+++ b/aten/src/ATen/core/function_schema.h
+@@ -206,7 +206,7 @@ inline bool operator!=(const Argument& lhs, const Argument& rhs) {
+   return !(lhs == rhs);
+ }
+ 
+-enum struct TORCH_API SchemaArgType { input, output };
+enum struct SchemaArgType { input, output };
+ 
+ /**
+  * struct SchemaArgument
+diff --git a/torch/csrc/jit/runtime/static/native_ops.cpp b/torch/csrc/jit/runtime/static/native_ops.cpp
+index 7f3cbcf..fe80ab1 100644
+--- a/torch/csrc/jit/runtime/static/native_ops.cpp
+++ b/torch/csrc/jit/runtime/static/native_ops.cpp
+@@ -1055,7 +1055,7 @@ namespace {
+   execution is completed, future is marked as complete to
+   indicate aten::wait() to proceed
+ */
+-class TORCH_API ForkedSubgraphSRLauncher {
+class ForkedSubgraphSRLauncher {
+  public:
+   ForkedSubgraphSRLauncher(
+       std::shared_ptr<StaticModule> smodule,
--- a/ports/libtorch/fix-build.patch
+++ b/ports/libtorch/fix-build.patch
@ -0,0 +1,184 @@
+diff --git a/c10/cuda/CMakeLists.txt b/c10/cuda/CMakeLists.txt
+index c0628d0c96..5dfb7f2abb 100644
+--- a/c10/cuda/CMakeLists.txt
+++ b/c10/cuda/CMakeLists.txt
+@@ -74,7 +74,11 @@ add_subdirectory(test)
+ # Note: for now, we will put all export path into one single Caffe2Targets group
+ # to deal with the cmake deployment need. Inside the Caffe2Targets set, the
+ # individual libraries like libc10.so and libcaffe2.so are still self-contained.
+-install(TARGETS c10_cuda EXPORT Caffe2Targets DESTINATION lib)
+install(TARGETS c10_cuda EXPORT Caffe2Targets 
+  RUNTIME DESTINATION bin
+  ARCHIVE DESTINATION lib
+  LIBRARY DESTINATION lib
+  )
+ foreach(file ${C10_CUDA_HEADERS})
+   get_filename_component( dir ${file} DIRECTORY )
+   install( FILES ${file} DESTINATION include/c10/cuda/${dir} )
+diff --git a/c10/util/Exception.h b/c10/util/Exception.h
+index ecbc7c1345..2dd1001c9e 100644
+--- a/c10/util/Exception.h
+++ b/c10/util/Exception.h
+@@ -14,6 +14,9 @@
+ #define __func__ __FUNCTION__
+ #endif
+ 
+#ifdef _MSC_VER
+#pragma warning(suppress : 4251)
+#endif
+ namespace c10 {
+ 
+ /// The primary ATen error class.
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 2c5c5b95d8..8c1742591f 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1581,19 +1581,10 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+       target_compile_options(onnx PRIVATE -Wno-deprecated-declarations)
+     endif()
+   else()
+-    add_library(onnx UNKNOWN IMPORTED)
+-    find_library(ONNX_LIBRARY NAMES onnx REQUIRED)
+-    if(NOT ONNX_LIBRARY)
+-      message(FATAL_ERROR "Cannot find onnx")
+-    endif()
+-    set_property(TARGET onnx PROPERTY IMPORTED_LOCATION ${ONNX_LIBRARY})
+-    add_library(onnx_proto UNKNOWN IMPORTED)
+-    find_library(ONNX_PROTO_LIBRARY NAMES onnx_proto REQUIRED)
+-    if(NOT ONNX_PROTO_LIBRARY)
+-      message(FATAL_ERROR "Cannot find onnx")
+-    endif()
+-    set_property(TARGET onnx_proto PROPERTY IMPORTED_LOCATION ${ONNX_PROTO_LIBRARY})
+-    message("-- Found onnx: ${ONNX_LIBRARY} ${ONNX_PROTO_LIBRARY}")
+    find_package(onnx REQUIRED)
+    set(ONNX_LIBRARY onnx)
+    set(ONNX_PROTO_LIBRARY onnx_proto)
+    message("-- Found onnx")
+     list(APPEND Caffe2_DEPENDENCY_LIBS onnx_proto onnx)
+   endif()
+   include_directories(${FOXI_INCLUDE_DIRS})
+diff --git a/pyproject.toml b/pyproject.toml
+index eb764cb895..c70f317c2f 100644
+--- a/pyproject.toml
+++ b/pyproject.toml
+@@ -4,11 +4,9 @@ requires = [
+     "wheel",
+     "astunparse",
+     "numpy",
+-    "ninja",
+     "pyyaml",
+-    "cmake",
+     "typing-extensions",
+-    "requests",
+#    "requests",
+ ]
+ # Use legacy backend to import local packages in setup.py
+ build-backend = "setuptools.build_meta:__legacy__"
+diff --git a/setup.py b/setup.py
+index 17bf16b89a..b90ebbc9eb 100644
+--- a/setup.py
+++ b/setup.py
+@@ -263,7 +263,7 @@ from tools.setup_helpers.env import build_type, IS_DARWIN, IS_LINUX, IS_WINDOWS
+ ################################################################################
+ 
+ VERBOSE_SCRIPT = True
+-RUN_BUILD_DEPS = True
+RUN_BUILD_DEPS = False
+ # see if the user passed a quiet flag to setup.py arguments and respect
+ # that in our parts of the build
+ EMIT_BUILD_WARNING = False
+@@ -307,9 +307,9 @@ else:
+ 
+ # Constant known variables used throughout this file
+ cwd = os.path.dirname(os.path.abspath(__file__))
+-lib_path = os.path.join(cwd, "torch", "lib")
+lib_path = os.path.join(cwd, "../../@TARGET_TRIPLET@/", "lib")
+ third_party_path = os.path.join(cwd, "third_party")
+-caffe2_build_dir = os.path.join(cwd, "build")
+caffe2_build_dir = os.path.join(cwd, "../../@TARGET_TRIPLET@")
+ 
+ # CMAKE: full path to python library
+ if IS_WINDOWS:
+@@ -759,7 +759,7 @@ class build_ext(setuptools.command.build_ext.build_ext):
+             with open(filename) as f:
+                 return json.load(f)
+ 
+-        ninja_files = glob.glob("build/*compile_commands.json")
+        ninja_files = glob.glob("../../@TARGET_TRIPLET@/*compile_commands.json")
+         cmake_files = glob.glob("torch/lib/build/*/compile_commands.json")
+         all_commands = [entry for f in ninja_files + cmake_files for entry in load(f)]
+ 
+diff --git a/third_party/nvfuser/csrc/instrumentation.h b/third_party/nvfuser/csrc/instrumentation.h
+index cd57825a24..75e81e9fac 100644
+--- a/third_party/nvfuser/csrc/instrumentation.h
+++ b/third_party/nvfuser/csrc/instrumentation.h
+@@ -2,7 +2,7 @@
+ 
+ #include <utils.h>
+ 
+-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
+ 
+ // NOLINTNEXTLINE(modernize-deprecated-headers)
+ #include <stdio.h>
+diff --git a/tools/setup_helpers/env.py b/tools/setup_helpers/env.py
+index d87e97a2bb..e1bc42e87c 100644
+--- a/tools/setup_helpers/env.py
+++ b/tools/setup_helpers/env.py
+@@ -19,7 +19,7 @@ CONDA_DIR = os.path.join(os.path.dirname(sys.executable), "..")
+ 
+ IS_64BIT = struct.calcsize("P") == 8
+ 
+-BUILD_DIR = "build"
+BUILD_DIR = "../../@TARGET_TRIPLET@"
+ 
+ 
+ def check_env_flag(name: str, default: str = "") -> bool:
+diff --git a/torch/lib/libshm/CMakeLists.txt b/torch/lib/libshm/CMakeLists.txt
+index 20158a9a25..7d3c475a9b 100644
+--- a/torch/lib/libshm/CMakeLists.txt
+++ b/torch/lib/libshm/CMakeLists.txt
+@@ -71,6 +71,10 @@ target_link_libraries(torch_shm_manager PRIVATE shm)
+ set_target_properties(torch_shm_manager PROPERTIES
+   INSTALL_RPATH "${_rpath_portable_origin}/../lib")
+ 
+-install(TARGETS shm LIBRARY DESTINATION ${LIBSHM_INSTALL_LIB_SUBDIR})
+install(TARGETS shm 
+  RUNTIME DESTINATION bin
+  ARCHIVE DESTINATION lib
+  LIBRARY DESTINATION lib
+)
+ install(FILES libshm.h DESTINATION "include")
+ install(TARGETS torch_shm_manager DESTINATION "bin")
+diff --git a/torch/lib/libshm_windows/CMakeLists.txt b/torch/lib/libshm_windows/CMakeLists.txt
+index df2a106493..b86e0b6872 100644
+--- a/torch/lib/libshm_windows/CMakeLists.txt
+++ b/torch/lib/libshm_windows/CMakeLists.txt
+@@ -19,9 +19,12 @@ target_include_directories(shm PRIVATE
+ target_link_libraries(shm torch c10)
+ 
+ 
+-install(TARGETS shm DESTINATION "${LIBSHM_INSTALL_LIB_SUBDIR}")
+install(TARGETS shm 
+  LIBRARY DESTINATION "${LIBSHM_INSTALL_LIB_SUBDIR}"
+  ARCHIVE DESTINATION "${LIBSHM_INSTALL_LIB_SUBDIR}"
+  RUNTIME DESTINATION "${LIBSHM_INSTALL_BIN_SUBDIR}")
+ install(FILES libshm.h DESTINATION "include")
+ 
+ if(MSVC AND BUILD_SHARED_LIBS)
+-  install(FILES $<TARGET_PDB_FILE:shm> DESTINATION "${LIBSHM_INSTALL_LIB_SUBDIR}" OPTIONAL)
+  install(FILES $<TARGET_PDB_FILE:shm> DESTINATION "${LIBSHM_INSTALL_BIN_SUBDIR}" OPTIONAL)
+ endif()
+diff --git a/torch/utils/cpp_extension.py b/torch/utils/cpp_extension.py
+index 6600eb3cb5..1d5e451c14 100644
+--- a/torch/utils/cpp_extension.py
+++ b/torch/utils/cpp_extension.py
+@@ -38,7 +38,7 @@ SHARED_FLAG = '/DLL' if IS_WINDOWS else '-shared'
+ 
+ _HERE = os.path.abspath(__file__)
+ _TORCH_PATH = os.path.dirname(os.path.dirname(_HERE))
+-TORCH_LIB_PATH = os.path.join(_TORCH_PATH, 'lib')
+TORCH_LIB_PATH = os.path.join(_TORCH_PATH, '../../../@TARGET_TRIPLET@', 'lib')
+ 
+ 
+ SUBPROCESS_DECODE_ARGS = ('oem',) if IS_WINDOWS else ()
--- a/ports/libtorch/fix-c10-glog.patch
+++ b/ports/libtorch/fix-c10-glog.patch
@ -1,61 +0,0 @@
-diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
-index fe74e49..06e09d6 100644
--- a/c10/util/Logging.cpp
-+++ b/c10/util/Logging.cpp
-@@ -4,7 +4,7 @@
- #ifdef FBCODE_CAFFE2
- #include <folly/synchronization/SanitizeThread.h>
- #endif
-
-+#include <glog/logging.h>
- #include <algorithm>
- #include <cstdlib>
- #include <cstring>
-@@ -192,23 +194,13 @@ C10_DEFINE_int(
-     google::GLOG_WARNING,
-     "The minimum log level that caffe2 will output.");
- 
-// Google glog's api does not have an external function that allows one to check
-// if glog is initialized or not. It does have an internal function - so we are
-// declaring it here. This is a hack but has been used by a bunch of others too
-// (e.g. Torch).
-namespace google {
-namespace glog_internal_namespace_ {
-bool IsGoogleLoggingInitialized();
-} // namespace glog_internal_namespace_
-} // namespace google
-
- namespace c10 {
- namespace {
- 
- void initGoogleLogging(char const* name) {
- #if !defined(_MSC_VER)
-   // This trick can only be used on UNIX platforms
-  if (!::google::glog_internal_namespace_::IsGoogleLoggingInitialized())
-+  if (!::google::IsGoogleLoggingInitialized())
- #endif
-   {
-     ::google::InitGoogleLogging(name);
-diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
-index 1864f17..b958fa0 100644
--- a/c10/util/Logging.cpp
-+++ b/c10/util/Logging.cpp
-@@ -154,18 +154,6 @@ void setLogLevelFlagFromEnv();
- } // namespace detail
- } // namespace c10
- 
-#if defined(C10_USE_GFLAGS) && defined(C10_USE_GLOG)
-// When GLOG depends on GFLAGS, these variables are being defined in GLOG
-// directly via the GFLAGS definition, so we will use DECLARE_* to declare
-// them, and use them in Caffe2.
-// GLOG's minloglevel
-DECLARE_int32(minloglevel);
-// GLOG's verbose log value.
-DECLARE_int32(v);
-// GLOG's logtostderr value
-DECLARE_bool(logtostderr);
-#endif // defined(C10_USE_GFLAGS) && defined(C10_USE_GLOG)
-
- #if !defined(C10_USE_GLOG)
- // This backward compatibility flags are in order to deal with cases where
- // Caffe2 are not built with glog, but some init flags still pass in these
--- a/ports/libtorch/fix-cmake.patch
+++ b/ports/libtorch/fix-cmake.patch
@ -1,142 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 4bf1e85..48b01ae 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -693,7 +693,7 @@ endif()
- 
- set(BUILD_ONEDNN_GRAPH OFF)
- 
-include(cmake/Dependencies.cmake)
-+include(cmake/vcpkg-dependencies.cmake)
- 
- if(USE_CUDA AND (CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 10.2) AND (CMAKE_HOST_SYSTEM_NAME MATCHES "Windows"))
-   # CUDA < 10.2 doesn't support compiling and extracting header dependencies in
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 48b01ae..8e46275 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -660,7 +660,7 @@ if(NOT CMAKE_BUILD_TYPE)
- endif()
- 
- # The below means we are cross compiling for arm64 or x86_64 on MacOSX
-if(NOT IOS AND CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64|arm64)$")
-+if(false)
-   set(CROSS_COMPILING_MACOSX TRUE)
-   # We need to compile a universal protoc to not fail protobuf build
-   # We set CMAKE_TRY_COMPILE_TARGET_TYPE to STATIC_LIBRARY (vs executable) to succeed the cmake compiler check for cross-compiling
-@@ -678,6 +678,7 @@ if(NOT IOS AND CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_OSX_ARCHITECTURES M
-   set(PROTOBUF_PROTOC_EXECUTABLE "${PROJECT_SOURCE_DIR}/build_host_protoc/bin/protoc")
-   set(CAFFE2_CUSTOM_PROTOC_EXECUTABLE "${PROJECT_SOURCE_DIR}/build_host_protoc/bin/protoc")
- endif()
-+include(cmake/ProtoBuf.cmake)
- 
- # ---[ Misc checks to cope with various compiler modes
- include(cmake/MiscCheck.cmake)
-diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
-index 328210c..c407442 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -112,13 +112,8 @@ endif()
- # addressed yet.
- 
- if(NOT MSVC AND USE_XNNPACK)
-  if(NOT TARGET fxdiv)
-    set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
-    set(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
-    add_subdirectory(
-      "${FXDIV_SOURCE_DIR}"
-      "${CMAKE_BINARY_DIR}/FXdiv")
-  endif()
-+  find_path(FXDIV_INCLUDE_DIRS "fxdiv.h" REQUIRED)
-+  include_directories(${FXDIV_INCLUDE_DIRS})
- endif()
- 
- add_subdirectory(core)
-diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
-index c407442..16fa898 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -1011,10 +1011,6 @@ if(USE_CUDA OR USE_ROCM)
-   include(${TORCH_SRC_DIR}/csrc/jit/codegen/cuda/nvfuser.cmake)
- endif()
- 
-if(NOT MSVC AND USE_XNNPACK)
-  TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
-endif()
-
- # ==========================================================
- # formerly-libtorch flags
- # ==========================================================
-diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
-index 41b1a1a..3b390ec 100644
--- a/c10/CMakeLists.txt
-+++ b/c10/CMakeLists.txt
-@@ -110,7 +110,11 @@ endif()
- # Note: for now, we will put all export path into one single Caffe2Targets group
- # to deal with the cmake deployment need. Inside the Caffe2Targets set, the
- # individual libraries like libc10.so and libcaffe2.so are still self-contained.
-install(TARGETS c10 EXPORT Caffe2Targets DESTINATION lib)
-+install(TARGETS c10 EXPORT Caffe2Targets
-+  RUNTIME DESTINATION bin
-+  LIBRARY DESTINATION lib
-+  ARCHIVE DESTINATION lib
-+)
- install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}
-         DESTINATION include
-         FILES_MATCHING PATTERN "*.h")
-diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
-index 16fa898..7a343f7 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -1547,7 +1547,11 @@ endif()
- 
- caffe2_interface_library(torch torch_library)
- 
-install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
-+install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets
-+  RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
-+  LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
-+  ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
-+)
- 
- if(USE_CUDA)
-   install(TARGETS torch_cuda torch_cuda_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
-@@ -1558,7 +1562,11 @@ if(USE_CUDA)
- elseif(USE_ROCM)
-   install(TARGETS torch_hip torch_hip_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
- endif()
-install(TARGETS torch torch_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
-+install(TARGETS torch torch_library EXPORT Caffe2Targets
-+  RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
-+  LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
-+  ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
-+)
- 
- target_link_libraries(torch PUBLIC torch_cpu_library)
- 
-@@ -1688,7 +1696,11 @@ if(BUILD_SHARED_LIBS)
-     target_link_libraries(torch_global_deps TBB::tbb)
-   endif()
- 
-  install(TARGETS torch_global_deps DESTINATION "${TORCH_INSTALL_LIB_DIR}")
-+  install(TARGETS torch_global_deps
-+    RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
-+    LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
-+    ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
-+  )
- endif()
- 
- # ---[ Caffe2 HIP sources.
-diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
-index 7a343f7..3390f9b 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -1673,7 +1673,7 @@ target_link_libraries(torch_cpu PRIVATE flatbuffers)
- # namespaces, so libtorch is loaded with all its dependencies in a local scope.
- # That usually leads to missing symbol errors at run-time, so to avoid a situation like
- # this we have to preload those libs in a global namespace.
-if(BUILD_SHARED_LIBS)
-+if(BUILD_SHARED_LIBS AND (NOT WIN32))
-   add_library(torch_global_deps SHARED ${TORCH_SRC_DIR}/csrc/empty.c)
-   if(HAVE_SOVERSION)
-     set_target_properties(torch_global_deps PROPERTIES
--- a/ports/libtorch/fix-fbgemm-include.patch
+++ b/ports/libtorch/fix-fbgemm-include.patch
@ -1,32 +0,0 @@
-diff --git a/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp b/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp
-index 224a66f..eb42317 100644
--- a/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp
-+++ b/aten/src/ATen/native/quantized/cpu/qembeddingbag_prepack.cpp
-@@ -8,7 +8,10 @@
- #include <torch/library.h>
- 
- #include <c10/util/irange.h>
-
-+#ifdef USE_FBGEMM
-+#include <fbgemm/Fbgemm.h>
-+#include <fbgemm/FbgemmEmbedding.h>
-+#endif
- int register_embedding_params();
- 
- /*
-diff --git a/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp b/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
-index f94d27b..299b88b 100644
--- a/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
-+++ b/aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
-@@ -4,7 +4,10 @@
- #include <ATen/native/quantized/cpu/fbgemm_utils.h>
- #include <c10/util/irange.h>
- #include <torch/library.h>
-
-+#ifdef USE_FBGEMM
-+#include <fbgemm/Fbgemm.h>
-+#include <fbgemm/FbgemmEmbedding.h>
-+#endif
- int register_embedding_params();
- 
- at::Tensor PackedEmbeddingBagWeight::unpack() {
--- a/ports/libtorch/fix-glog.patch
+++ b/ports/libtorch/fix-glog.patch
@ -0,0 +1,30 @@
+diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
+index ff8e1d6cc..306dac51f 100644
+--- a/c10/util/Logging.cpp	
+++ b/c10/util/Logging.cpp
+@@ -218,24 +218,11 @@ C10_DEFINE_int(
+     google::GLOG_WARNING,
+     "The minimum log level that caffe2 will output.");
+ 
+-// Google glog's api does not have an external function that allows one to check
+-// if glog is initialized or not. It does have an internal function - so we are
+-// declaring it here. This is a hack but has been used by a bunch of others too
+-// (e.g. Torch).
+-namespace google {
+-namespace glog_internal_namespace_ {
+-bool IsGoogleLoggingInitialized();
+-} // namespace glog_internal_namespace_
+-} // namespace google
+-
+ namespace c10 {
+ namespace {
+ 
+ void initGoogleLogging(char const* name) {
+-#if !defined(_MSC_VER)
+-  // This trick can only be used on UNIX platforms
+-  if (!::google::glog_internal_namespace_::IsGoogleLoggingInitialized())
+-#endif
+  if (!::google::IsGoogleLoggingInitialized())
+   {
+     ::google::InitGoogleLogging(name);
+ #if !defined(_MSC_VER)
--- a/ports/libtorch/pytorch-pr-85958.patch
+++ b/ports/libtorch/pytorch-pr-85958.patch
@ -1,15 +1,15 @@
 diff --git a/aten/src/ATen/cpu/vec/vec_base.h b/aten/src/ATen/cpu/vec/vec_base.h
-index 635ec8c82e5d..1974d55943b3 100644
--- a/aten/src/ATen/cpu/vec/vec_base.h
+index f68e9c6eb..c42a9b8bb 100644
+--- a/aten/src/ATen/cpu/vec/vec_base.h	
 +++ b/aten/src/ATen/cpu/vec/vec_base.h
-@@ -132,8 +132,9 @@ struct Vectorized {
+@@ -150,9 +150,8 @@ public:
   // versions GCC/Clang have buggy determinations on whether or not an
   // identifier is odr-used or not, and in any case it's hard to tell if
   // a variable is odr-used or not.  So best to just cut the problem at the root.
-+  static constexpr size_type size_T = sizeof(T);  // Workaround to compile with VS2022.
+-  static constexpr size_type size_T = sizeof(T);  // Workaround to compile with VS2022.
   static constexpr size_type size() {
-    return VECTOR_WIDTH / sizeof(T);
-+    return VECTOR_WIDTH / size_T;
+-    return VECTOR_WIDTH / size_T;
+    return VECTOR_WIDTH / sizeof(T);
   }
   Vectorized() : values{static_cast<T>(0)} {}
   Vectorized(T val) {
--- a/ports/libtorch/fix-onnx-case.patch
+++ b/ports/libtorch/fix-onnx-case.patch
@ -0,0 +1,13 @@
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 8c1742591..a2ca73b6e 100644
+--- a/cmake/Dependencies.cmake	
+++ b/cmake/Dependencies.cmake
+@@ -1581,7 +1581,7 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+       target_compile_options(onnx PRIVATE -Wno-deprecated-declarations)
+     endif()
+   else()
+-    find_package(onnx REQUIRED)
+    find_package(ONNX REQUIRED)
+     set(ONNX_LIBRARY onnx)
+     set(ONNX_PROTO_LIBRARY onnx_proto)
+     message("-- Found onnx")
--- a/ports/libtorch/fix-sleef.patch
+++ b/ports/libtorch/fix-sleef.patch
@ -0,0 +1,22 @@
+diff --git a/aten/src/ATen/CMakeLists.txt b/aten/src/ATen/CMakeLists.txt
+index f3281d3cb..c8b6ed30e 100644
+--- a/aten/src/ATen/CMakeLists.txt	
+++ b/aten/src/ATen/CMakeLists.txt
+@@ -418,7 +418,7 @@ if(NOT MSVC AND NOT EMSCRIPTEN AND NOT INTERN_BUILD_MOBILE)
+     list(APPEND ATen_THIRD_PARTY_INCLUDE ${CMAKE_BINARY_DIR}/include)
+     link_directories(${CMAKE_BINARY_DIR}/sleef/lib)
+   else()
+-    add_library(sleef SHARED IMPORTED)
+    add_library(sleef UNKNOWN IMPORTED)
+     find_library(SLEEF_LIBRARY sleef)
+     if(NOT SLEEF_LIBRARY)
+       message(FATAL_ERROR "Cannot find sleef")
+@@ -426,7 +426,7 @@ if(NOT MSVC AND NOT EMSCRIPTEN AND NOT INTERN_BUILD_MOBILE)
+     message("Found sleef: ${SLEEF_LIBRARY}")
+     set_target_properties(sleef PROPERTIES IMPORTED_LOCATION "${SLEEF_LIBRARY}")
+   endif()
+-  list(APPEND ATen_CPU_DEPENDENCY_LIBS sleef)
+  list(APPEND ATen_CPU_DEPENDENCY_LIBS ${SLEEF_LIBRARY})
+ 
+   set(CMAKE_C_FLAGS_DEBUG ${OLD_CMAKE_C_FLAGS_DEBUG})
+   set(CMAKE_CXX_FLAGS ${OLD_CMAKE_CXX_FLAGS})
--- a/ports/libtorch/fix-windows.patch
+++ b/ports/libtorch/fix-windows.patch
@ -1,100 +0,0 @@
-diff --git a/torch/csrc/jit/tensorexpr/bounds_overlap.h b/torch/csrc/jit/tensorexpr/bounds_overlap.h
-index c78884d..baf44bb 100644
--- a/torch/csrc/jit/tensorexpr/bounds_overlap.h
-+++ b/torch/csrc/jit/tensorexpr/bounds_overlap.h
-@@ -65,6 +65,10 @@ enum OverlapKind { ContainedOrEqual, Contains, PartialOverlap, NoOverlap };
- //     operator
- //     NOT_DETERMINED: Some elements satify the given comparison operator and
- //     some elements not
-+#if defined(_WIN32)
-+#undef TRUE
-+#undef FALSE
-+#endif
- enum CmpEvalResult { TRUE, FALSE, NOT_DETERMINED };
- 
- // Returns the kind of overlap between Bound A and Bound A in a single
-diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
-index 3390f9b..578d0bd 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -1425,7 +1425,7 @@ target_include_directories(torch_cpu SYSTEM PRIVATE "${Caffe2_DEPENDENCY_INCLUDE
- torch_set_target_props(torch_cpu)
- 
- 
-target_compile_options(torch_cpu PRIVATE "-DCAFFE2_BUILD_MAIN_LIB")
-+target_compile_definitions(torch_cpu PRIVATE CAFFE2_BUILD_MAIN_LIB)
- if(BUILD_SPLIT_CUDA)
-   target_compile_options(torch_cuda_cu PRIVATE "-DTORCH_CUDA_CU_BUILD_MAIN_LIB")
-   target_compile_options(torch_cuda_cpp PRIVATE "-DTORCH_CUDA_CPP_BUILD_MAIN_LIB")
-@@ -1525,12 +1525,12 @@ if(MSVC AND NOT BUILD_SHARED_LIBS)
-   #
-   # NB: This must be target_compile_definitions, not target_compile_options,
-   # as the latter is not respected by nvcc
-  target_compile_definitions(torch_cpu PUBLIC "AT_CORE_STATIC_WINDOWS=1")
-+  target_compile_definitions(torch_cpu PUBLIC AT_CORE_STATIC_WINDOWS=1)
- endif()
- if(MSVC AND BUILD_SHARED_LIBS)
-   # ONNX is linked statically and needs to be exported from this library
-   # to be used externally. Make sure that references match the export.
-  target_compile_options(torch_cpu PRIVATE "-DONNX_BUILD_MAIN_LIB")
-+  target_compile_definitions(torch_cpu PRIVATE ONNX_BUILD_MAIN_LIB)
- endif()
- 
- caffe2_interface_library(torch_cpu torch_cpu_library)
-diff --git a/aten/src/ATen/Parallel.h b/aten/src/ATen/Parallel.h
-index 0b494d9..06b5d88 100644
--- a/aten/src/ATen/Parallel.h
-+++ b/aten/src/ATen/Parallel.h
-@@ -28,13 +28,7 @@ TORCH_API bool in_parallel_region();
- namespace internal {
- 
- // Initialise num_threads lazily at first parallel call
-inline TORCH_API void lazy_init_num_threads() {
-  thread_local bool init = false;
-  if (C10_UNLIKELY(!init)) {
-    at::init_num_threads();
-    init = true;
-  }
-}
-+TORCH_API void lazy_init_num_threads();
- 
- TORCH_API void set_thread_num(int);
- 
-diff --git a/aten/src/ATen/ParallelCommon.cpp b/aten/src/ATen/ParallelCommon.cpp
-index 068c9d3..b869fde 100644
--- a/aten/src/ATen/ParallelCommon.cpp
-+++ b/aten/src/ATen/ParallelCommon.cpp
-@@ -17,6 +17,17 @@
- 
- namespace at {
- 
-+namespace internal {
-+/// @see https://github.com/pytorch/pytorch/issues/87957
-+void lazy_init_num_threads() {
-+  thread_local bool init = false;
-+  if (C10_UNLIKELY(!init)) {
-+    at::init_num_threads();
-+    init = true;
-+  }
-+}
-+} // namespace internal
-+
- namespace {
- 
- const char* get_env_var(
-diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
-index 3b390ec..7cfa2b0 100644
--- a/c10/CMakeLists.txt
-+++ b/c10/CMakeLists.txt
-@@ -45,7 +45,10 @@ if(HAVE_SOVERSION)
-       VERSION ${TORCH_VERSION} SOVERSION ${TORCH_SOVERSION})
- endif()
- # If building shared library, set dllimport/dllexport proper.
-target_compile_options(c10 PRIVATE "-DC10_BUILD_MAIN_LIB")
-+target_compile_definitions(c10 PRIVATE C10_BUILD_MAIN_LIB)
-+if(WIN32)
-+  target_compile_definitions(c10 PRIVATE NOMINMAX)
-+endif()
- # Enable hidden visibility if compiler supports it.
- if(${COMPILER_SUPPORTS_HIDDEN_VISIBILITY})
-   target_compile_options(c10 PRIVATE "-fvisibility=hidden")
--- a/ports/libtorch/fix_werror.patch
+++ b/ports/libtorch/fix_werror.patch
@ -1,12 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 8e46275..b505ec5 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -788,7 +788,6 @@ if(NOT MSVC)
-   # Details at http://eigen.tuxfamily.org/bz/show_bug.cgi?id=1459
-   string(APPEND CMAKE_CXX_FLAGS " -Wall")
-   string(APPEND CMAKE_CXX_FLAGS " -Wextra")
-  string(APPEND CMAKE_CXX_FLAGS " -Werror=return-type")
-   string(APPEND CMAKE_CXX_FLAGS " -Wno-missing-field-initializers")
-   string(APPEND CMAKE_CXX_FLAGS " -Wno-type-limits")
-   string(APPEND CMAKE_CXX_FLAGS " -Wno-array-bounds")
--- a/ports/libtorch/fxdiv.patch
+++ b/ports/libtorch/fxdiv.patch
@ -0,0 +1,44 @@
+diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+index d18523352..1f53945cf 100644
+--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+@@ -389,12 +389,12 @@ if(NOT TARGET fxdiv AND NOT USE_SYSTEM_FXDIV)
+     "${FXDIV_SOURCE_DIR}"
+     "${CONFU_DEPENDENCIES_BINARY_DIR}/fxdiv")
+ elseif(NOT TARGET fxdiv AND USE_SYSTEM_FXDIV)
+-  find_file(FXDIV_HDR fxdiv.h PATH_SUFFIXES include)
+  find_path(FXDIV_HDR fxdiv.h PATH_SUFFIXES include)
+   if(NOT FXDIV_HDR)
+     message(FATAL_ERROR "Cannot find fxdiv")
+   endif()
+-  add_library(fxdiv STATIC "${FXDIV_HDR}")
+-  set_property(TARGET fxdiv PROPERTY LINKER_LANGUAGE C)
+  add_library(fxdiv INTERFACE IMPORTED)
+  target_include_directories(fxdiv INTERFACE "${FXDIV_HDR}")
+ endif()
+ target_link_libraries(pytorch_qnnpack PRIVATE fxdiv)
+ 
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index f7a595e1b..fb315c435 100644
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -108,11 +108,18 @@ endif()
+ # addressed yet.
+ 
+ if(NOT MSVC AND USE_XNNPACK)
+-  if(NOT TARGET fxdiv)
+  if(NOT TARGET fxdiv AND NOT USE_SYSTEM_FXDIV)
+     set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+     set(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+     add_subdirectory(
+       "${FXDIV_SOURCE_DIR}"
+       "${CMAKE_BINARY_DIR}/FXdiv")
+  elseif(NOT TARGET fxdiv AND USE_SYSTEM_FXDIV)
+    find_path(FXDIV_HDR fxdiv.h PATH_SUFFIXES include)
+    if(NOT FXDIV_HDR)
+      message(FATAL_ERROR "Cannot find fxdiv")
+    endif()
+    add_library(fxdiv INTERFACE IMPORTED)
+    target_include_directories(fxdiv INTERFACE "${FXDIV_HDR}")
+   endif()
+ endif()
--- a/ports/libtorch/kineto.patch
+++ b/ports/libtorch/kineto.patch
@ -0,0 +1,34 @@
+diff --git a/libkineto/CMakeLists.txt b/libkineto/CMakeLists.txt
+index 7d36ffb9d4..8f97998114 100644
+--- a/libkineto/CMakeLists.txt
+++ b/libkineto/CMakeLists.txt
+@@ -111,27 +111,8 @@ endif()
+ target_compile_options(kineto_base PRIVATE "${KINETO_COMPILE_OPTIONS}")
+ target_compile_options(kineto_api PRIVATE "${KINETO_COMPILE_OPTIONS}")
+ 
+-if(NOT TARGET fmt)
+-  if(NOT FMT_SOURCE_DIR)
+-    set(FMT_SOURCE_DIR "${LIBKINETO_THIRDPARTY_DIR}/fmt"
+-      CACHE STRING "fmt source directory from submodules")
+-  endif()
+-
+-  # Build FMT.
+-  # FMT and some other libraries use BUILD_SHARED_LIBS to control
+-  # the library type.
+-  # Save and restore the value after configuring FMT
+-  set(TEMP_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+-  set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build shared libs" FORCE)
+-  set(FMT_LIBRARY_TYPE static CACHE STRING "Set lib type to static")
+-  add_subdirectory("${FMT_SOURCE_DIR}" "${LIBKINETO_BINARY_DIR}/fmt")
+-  set_property(TARGET fmt PROPERTY POSITION_INDEPENDENT_CODE ON)
+-  set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
+-endif()
+-
+-set(FMT_INCLUDE_DIR "${FMT_SOURCE_DIR}/include")
+-message(STATUS "Kineto: FMT_SOURCE_DIR = ${FMT_SOURCE_DIR}")
+-message(STATUS "Kineto: FMT_INCLUDE_DIR = ${FMT_INCLUDE_DIR}")
+find_package(fmt REQUIRED)
+
+ if (NOT CUPTI_INCLUDE_DIR)
+     set(CUPTI_INCLUDE_DIR "${CUDA_SOURCE_DIR}/extras/CUPTI/include")
+ endif()
--- a/ports/libtorch/missing_include.patch
+++ b/ports/libtorch/missing_include.patch
@ -1,57 +0,0 @@
-diff --git a/aten/src/ATen/core/dynamic_type.h b/aten/src/ATen/core/dynamic_type.h
-index a84644ddde0..b1b6b2c9ecc 100644
--- a/aten/src/ATen/core/dynamic_type.h
-+++ b/aten/src/ATen/core/dynamic_type.h
-@@ -1,5 +1,6 @@
- #pragma once
- 
-+#include <cstdint>
- #include <memory>
- #include <type_traits>
- 
-diff --git a/c10/util/BFloat16.h b/c10/util/BFloat16.h
-index 1ada02bba1c..f6f9fc65e0e 100644
--- a/c10/util/BFloat16.h
-+++ b/c10/util/BFloat16.h
-@@ -3,6 +3,7 @@
- // Defines the bloat16 type (brain floating-point). This representation uses
- // 1 bit for the sign, 8 bits for the exponent and 7 bits for the mantissa.
- 
-+#include <cstdint>
- #include <c10/macros/Macros.h>
- #include <cmath>
- #include <cstring>
-diff --git a/torch/csrc/jit/passes/quantization/quantization_type.h b/torch/csrc/jit/passes/quantization/quantization_type.h
-index ea5ca10b15a..c704f9c3037 100644
--- a/torch/csrc/jit/passes/quantization/quantization_type.h
-+++ b/torch/csrc/jit/passes/quantization/quantization_type.h
-@@ -1,5 +1,6 @@
- #pragma once
- #include <ostream>
-+#include <cstdint>
- 
- namespace torch {
- namespace jit {
-diff --git a/torch/csrc/jit/runtime/logging.cpp b/torch/csrc/jit/runtime/logging.cpp
-index 66712990368..b326f587bc8 100644
--- a/torch/csrc/jit/runtime/logging.cpp
-+++ b/torch/csrc/jit/runtime/logging.cpp
-@@ -1,6 +1,7 @@
- #include <torch/csrc/jit/runtime/logging.h>
- 
- #include <atomic>
-+#include <stdexcept>
- #include <mutex>
- #include <unordered_map>
- 
-diff --git a/torch/csrc/lazy/core/multi_wait.cpp b/torch/csrc/lazy/core/multi_wait.cpp
-index 6b9933518e0..db0742b32a3 100644
--- a/torch/csrc/lazy/core/multi_wait.cpp
-+++ b/torch/csrc/lazy/core/multi_wait.cpp
-@@ -1,5 +1,6 @@
- #include <torch/csrc/lazy/core/multi_wait.h>
- 
-+#include <stdexcept>
- #include <chrono>
- #include <exception>
- 
--- a/ports/libtorch/more-fixes.patch
+++ b/ports/libtorch/more-fixes.patch
@ -0,0 +1,351 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 7b8bc7cb6e..fa6da328b8 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -383,7 +383,7 @@ if(WIN32)
+       NAMES uv libuv
+       HINTS $ENV{CONDA_PREFIX}\\Library $ENV{PREFIX}\\Library
+       PATH_SUFFIXES lib
+-      NO_DEFAULT_PATH)
+      )
+     if(NOT libuv_tmp_LIBRARY)
+       set(USE_DISTRIBUTED OFF)
+       set(USE_GLOO OFF)
+@@ -1151,10 +1151,6 @@ if(BUILD_SHARED_LIBS)
+       ${PROJECT_SOURCE_DIR}/cmake/Modules_CUDA_fix
+       DESTINATION share/cmake/Caffe2/
+       COMPONENT dev)
+-  install(FILES
+-      ${PROJECT_SOURCE_DIR}/cmake/Modules/FindCUDAToolkit.cmake
+-      DESTINATION share/cmake/Caffe2/
+-      COMPONENT dev)
+   install(FILES
+       ${PROJECT_SOURCE_DIR}/cmake/Modules/FindCUSPARSELT.cmake
+       DESTINATION share/cmake/Caffe2/
+diff --git a/aten/src/ATen/native/vulkan/api/Allocator.h b/aten/src/ATen/native/vulkan/api/Allocator.h
+index d0c8bdf9ec..50df97ae09 100644
+--- a/aten/src/ATen/native/vulkan/api/Allocator.h
+++ b/aten/src/ATen/native/vulkan/api/Allocator.h
+@@ -47,7 +47,7 @@
+ #pragma clang diagnostic ignored "-Wunused-variable"
+ #endif /* __clang__ */
+ 
+-#include <include/vk_mem_alloc.h>
+#include <vk_mem_alloc.h>
+ 
+ #ifdef __clang__
+ #pragma clang diagnostic pop
+diff --git a/aten/src/ATen/native/vulkan/ops/Clamp.cpp b/aten/src/ATen/native/vulkan/ops/Clamp.cpp
+index dc22b987ef..f91fbf6788 100644
+--- a/aten/src/ATen/native/vulkan/ops/Clamp.cpp
+++ b/aten/src/ATen/native/vulkan/ops/Clamp.cpp
+@@ -1,3 +1,6 @@
+#define _USE_MATH_DEFINES 
+#include <cmath>
+
+ #include <ATen/native/vulkan/ops/Common.h>
+ #include <torch/library.h>
+ 
+diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
+index ad407e18cb..bbca41bab2 100644
+--- a/c10/CMakeLists.txt
+++ b/c10/CMakeLists.txt
+@@ -139,7 +139,11 @@ endif()
+ # Note: for now, we will put all export path into one single Caffe2Targets group
+ # to deal with the cmake deployment need. Inside the Caffe2Targets set, the
+ # individual libraries like libc10.so and libcaffe2.so are still self-contained.
+-install(TARGETS c10 EXPORT Caffe2Targets DESTINATION lib)
+install(TARGETS c10 EXPORT Caffe2Targets
+        RUNTIME DESTINATION bin
+        ARCHIVE DESTINATION lib
+        ARCHIVE DESTINATION lib
+        )
+ install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}
+         DESTINATION include
+         FILES_MATCHING PATTERN "*.h")
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index 74d0d55719..cf349c2fa8 100644
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -632,7 +632,11 @@ if(USE_CUDA)
+ 
+   target_link_libraries(caffe2_nvrtc ${CUDA_NVRTC} ${CUDA_CUDA_LIB} ${CUDA_NVRTC_LIB} ${DELAY_LOAD_FLAGS})
+   target_include_directories(caffe2_nvrtc PRIVATE ${CUDA_INCLUDE_DIRS})
+-  install(TARGETS caffe2_nvrtc DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS caffe2_nvrtc 
+    RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+    LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+    ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+  )
+   if(USE_NCCL)
+     list(APPEND Caffe2_GPU_SRCS
+       ${TORCH_SRC_DIR}/csrc/cuda/nccl.cpp)
+@@ -694,7 +698,11 @@ if(USE_ROCM)
+   add_library(caffe2_nvrtc SHARED ${ATen_NVRTC_STUB_SRCS})
+   target_link_libraries(caffe2_nvrtc ${PYTORCH_HIP_HCC_LIBRARIES} ${ROCM_HIPRTC_LIB})
+   target_compile_definitions(caffe2_nvrtc PRIVATE USE_ROCM __HIP_PLATFORM_HCC__)
+-  install(TARGETS caffe2_nvrtc DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS caffe2_nvrtc
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+    )
+ endif()
+ 
+ if(NOT NO_API AND NOT BUILD_LITE_INTERPRETER)
+@@ -1013,7 +1021,11 @@ elseif(USE_CUDA)
+           CUDA::culibos ${CMAKE_DL_LIBS})
+     endif()
+     set_source_files_properties(${CMAKE_CURRENT_SOURCE_DIR}/../aten/src/ATen/native/cuda/LinearAlgebraStubs.cpp PROPERTIES COMPILE_FLAGS "-DBUILD_LAZY_CUDA_LINALG")
+-    install(TARGETS torch_cuda_linalg DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+    install(TARGETS torch_cuda_linalg
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+    )
+   endif()
+ 
+   if(USE_PRECOMPILED_HEADERS)
+@@ -1475,14 +1487,26 @@ endif()
+ 
+ caffe2_interface_library(torch torch_library)
+ 
+-install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ 
+ if(USE_CUDA)
+-  install(TARGETS torch_cuda torch_cuda_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS torch_cuda torch_cuda_library EXPORT Caffe2Targets
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ elseif(USE_ROCM)
+-  install(TARGETS torch_hip torch_hip_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS torch_hip torch_hip_library EXPORT Caffe2Targets
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ endif()
+-install(TARGETS torch torch_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+install(TARGETS torch torch_library EXPORT Caffe2Targets
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ 
+ target_link_libraries(torch PUBLIC torch_cpu_library)
+ 
+@@ -1499,11 +1523,11 @@ endif()
+ 
+ # Install PDB files for MSVC builds
+ if(MSVC AND BUILD_SHARED_LIBS)
+-  install(FILES $<TARGET_PDB_FILE:torch_cpu> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
+  install(FILES $<TARGET_PDB_FILE:torch_cpu> DESTINATION "${TORCH_INSTALL_BIN_DIR}" OPTIONAL)
+   if(USE_CUDA)
+-    install(FILES $<TARGET_PDB_FILE:torch_cuda> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
+    install(FILES $<TARGET_PDB_FILE:torch_cuda> DESTINATION "${TORCH_INSTALL_BIN_DIR}" OPTIONAL)
+   elseif(USE_ROCM)
+-    install(FILES $<TARGET_PDB_FILE:torch_hip> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
+    install(FILES $<TARGET_PDB_FILE:torch_hip> DESTINATION "${TORCH_INSTALL_BIN_DIR}" OPTIONAL)
+   endif()
+ endif()
+ 
+@@ -1577,7 +1601,10 @@ if(BUILD_SHARED_LIBS)
+     target_link_libraries(torch_global_deps TBB::tbb)
+   endif()
+ 
+-  install(TARGETS torch_global_deps DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS torch_global_deps
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ endif()
+ 
+ # ---[ Caffe2 HIP sources.
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index cd27c728ff..2c5c5b95d8 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1130,7 +1130,7 @@ if(BUILD_PYTHON)
+     # Observers are required in the python build
+     caffe2_update_option(USE_OBSERVERS ON)
+   else()
+-    message(WARNING "Python dependencies not met. Not compiling with python. Suppress this warning with -DBUILD_PYTHON=OFF")
+    message(FATAL_ERROR "Python dependencies not met. Not compiling with python. Suppress this warning with -DBUILD_PYTHON=OFF")
+     caffe2_update_option(BUILD_PYTHON OFF)
+   endif()
+ endif()
+@@ -1234,7 +1234,7 @@ endif()
+ # ---[ LLVM
+ if(USE_LLVM)
+   message(STATUS "Looking for LLVM in ${USE_LLVM}")
+-  find_package(LLVM PATHS ${USE_LLVM} NO_DEFAULT_PATH)
+  find_package(LLVM)
+ 
+   if(LLVM_FOUND)
+     message(STATUS "Found LLVM ${LLVM_PACKAGE_VERSION}")
+@@ -1248,6 +1248,8 @@ endif(USE_LLVM)
+ # ---[ cuDNN
+ if(USE_CUDNN)
+   find_package(CUDNN REQUIRED)
+  set(CUDNN_FRONTEND_INCLUDE_DIR ${CMAKE_CURRENT_LIST_DIR}/../third_party/cudnn_frontend/include)
+  target_include_directories(torch::cudnn INTERFACE ${CUDNN_FRONTEND_INCLUDE_DIR})
+   target_include_directories(torch::cudnn INTERFACE "${CUDNN_INCLUDE_DIR}")
+ endif()
+ 
+diff --git a/cmake/Modules/FindCUSPARSELT.cmake b/cmake/Modules/FindCUSPARSELT.cmake
+index 8727f44a90..1e41281b1b 100644
+--- a/cmake/Modules/FindCUSPARSELT.cmake
+++ b/cmake/Modules/FindCUSPARSELT.cmake
+@@ -24,13 +24,13 @@ list(APPEND CMAKE_PREFIX_PATH ${CUSPARSELT_ROOT})
+ set(CUSPARSELT_INCLUDE_DIR $ENV{CUSPARSELT_INCLUDE_DIR} CACHE PATH "Folder containing NVIDIA cuSPARSELt header files")
+ 
+ find_path(CUSPARSELT_INCLUDE_PATH cusparseLt.h
+-  HINTS ${CUSPARSELT_INCLUDE_DIR}
+  HINTS ${CUSPARSELT_INCLUDE_DIR} ENV CUDA_PATH
+   PATH_SUFFIXES cuda/include cuda include)
+ 
+ set(CUSPARSELT_LIBRARY $ENV{CUSPARSELT_LIBRARY} CACHE PATH "Path to the cusparselt library file (e.g., libcusparseLt.so)")
+ 
+-find_library(CUSPARSELT_LIBRARY_PATH libcusparseLt.so
+-  PATHS ${CUSPARSELT_LIBRARY}
+find_library(CUSPARSELT_LIBRARY_PATH cusparseLt
+  PATHS ${CUSPARSELT_LIBRARY} ENV CUDA_PATH
+   PATH_SUFFIXES lib lib64 cuda/lib cuda/lib64 lib/x64)
+ 
+ find_package_handle_standard_args(CUSPARSELT DEFAULT_MSG CUSPARSELT_LIBRARY_PATH CUSPARSELT_INCLUDE_PATH)
+diff --git a/cmake/Modules/FindMAGMA.cmake b/cmake/Modules/FindMAGMA.cmake
+index d2bff53d9b..3af9012503 100644
+--- a/cmake/Modules/FindMAGMA.cmake
+++ b/cmake/Modules/FindMAGMA.cmake
+@@ -30,18 +30,13 @@ FIND_PATH(MAGMA_INCLUDE_DIR magma.h
+   HINTS $ENV{MAGMA_HOME} /usr/local/magma
+   PATH_SUFFIXES include)
+ 
+-IF (MAGMA_LIBRARIES)
+-  SET(MAGMA_FOUND TRUE)
+-ELSE (MAGMA_LIBRARIES)
+-  SET(MAGMA_FOUND FALSE)
+-ENDIF (MAGMA_LIBRARIES)
+-
+-add_library(torch::magma INTERFACE IMPORTED)
+-set_property(TARGET torch::magma
+-             PROPERTY INTERFACE_INCLUDE_DIRECTORIES "${MAGMA_INCLUDE_DIR}")
+-set_property(TARGET torch::magma
+-             PROPERTY INTERFACE_LINK_LIBRARIES "${MAGMA_LIBRARIES}")
+-
+if(NOT TARGET torch::magma)
+  add_library(torch::magma INTERFACE IMPORTED)
+  set_property(TARGET torch::magma
+               PROPERTY INTERFACE_INCLUDE_DIRECTORIES "${MAGMA_INCLUDE_DIR}")
+  set_property(TARGET torch::magma
+               PROPERTY INTERFACE_LINK_LIBRARIES "${MAGMA_LIBRARIES}")
+endif()
+ # Check for Magma V2
+ include(CheckPrototypeDefinition)
+ check_prototype_definition(magma_get_sgeqrf_nb
+@@ -53,3 +48,5 @@ if(MAGMA_V2)
+   set_property(TARGET torch::magma
+                PROPERTY INTERFACE_COMPILE_DEFINITIONS "MAGMA_V2")
+ endif(MAGMA_V2)
+
+find_package_handle_standard_args(MAGMA REQUIRED_VARS MAGMA_INCLUDE_DIR MAGMA_LIBRARIES)
+\ No newline at end of file
+diff --git a/test/cpp/jit/CMakeLists.txt b/test/cpp/jit/CMakeLists.txt
+index 2376f1bc43..a8ae4367fd 100644
+--- a/test/cpp/jit/CMakeLists.txt
+++ b/test/cpp/jit/CMakeLists.txt
+@@ -124,7 +124,7 @@ endif()
+ set(JIT_TEST_DEPENDENCIES torch gtest jitbackend_test backend_with_compiler)
+ 
+ if(MSVC)
+-  list(APPEND JIT_TEST_DEPENDENCIES onnx_library)
+  list(APPEND JIT_TEST_DEPENDENCIES onnx)
+ endif(MSVC)
+ 
+ target_link_libraries(test_jit PRIVATE ${JIT_TEST_DEPENDENCIES})
+diff --git a/third_party/nvfuser/CMakeLists.txt b/third_party/nvfuser/CMakeLists.txt
+index b148418af0..7b96083893 100644
+--- a/third_party/nvfuser/CMakeLists.txt
+++ b/third_party/nvfuser/CMakeLists.txt
+@@ -158,7 +158,10 @@ endif()
+ target_include_directories(${NVFUSER_CODEGEN}
+                            PUBLIC $<BUILD_INTERFACE:${NVFUSER_SRCS_DIR}>)
+ set_property(TARGET ${NVFUSER_CODEGEN} PROPERTY CXX_STANDARD 17)
+-install(TARGETS ${NVFUSER_CODEGEN} EXPORT NvfuserTargets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+install(TARGETS ${NVFUSER_CODEGEN} EXPORT NvfuserTargets
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ # installing nvfuser python tests
+ install(DIRECTORY "${NVFUSER_ROOT}/python_tests/"
+         DESTINATION "${TORCH_ROOT}/test/_nvfuser"
+diff --git a/torch/CMakeLists.txt b/torch/CMakeLists.txt
+index 62ee4c12a9..f75073cf3c 100644
+--- a/torch/CMakeLists.txt
+++ b/torch/CMakeLists.txt
+@@ -107,7 +107,7 @@ set(TORCH_PYTHON_LINK_FLAGS "")
+ 
+ if(MSVC)
+     string(APPEND TORCH_PYTHON_LINK_FLAGS " /NODEFAULTLIB:LIBCMT.LIB")
+-    list(APPEND TORCH_PYTHON_LINK_LIBRARIES ${PYTHON_LIBRARIES} onnx_library)
+    list(APPEND TORCH_PYTHON_LINK_LIBRARIES ${PYTHON_LIBRARIES} onnx)
+     if(NOT CMAKE_BUILD_TYPE MATCHES "Release")
+       string(APPEND TORCH_PYTHON_LINK_FLAGS " /DEBUG:FULL")
+     endif()
+@@ -374,7 +374,10 @@ if(NOT TORCH_PYTHON_LINK_FLAGS STREQUAL "")
+     set_target_properties(torch_python PROPERTIES LINK_FLAGS ${TORCH_PYTHON_LINK_FLAGS})
+ endif()
+ 
+-install(TARGETS torch_python DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+install(TARGETS torch_python       
+      RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}"
+      LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}"
+      ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ 
+ # Generate torch/version.py from the appropriate CMake cache variables.
+ if(${CMAKE_BUILD_TYPE} STREQUAL "Debug")
+diff --git a/torch/csrc/cuda/shared/nvtx.cpp b/torch/csrc/cuda/shared/nvtx.cpp
+index 4fb72c5f79..1abe1b76d6 100644
+--- a/torch/csrc/cuda/shared/nvtx.cpp
+++ b/torch/csrc/cuda/shared/nvtx.cpp
+@@ -1,7 +1,7 @@
+ #ifdef _WIN32
+ #include <wchar.h> // _wgetenv for nvtx
+ #endif
+-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
+ #include <torch/csrc/utils/pybind.h>
+ 
+ namespace torch::cuda::shared {
+diff --git a/torch/csrc/jit/tensorexpr/llvm_codegen.h b/torch/csrc/jit/tensorexpr/llvm_codegen.h
+index 7ab506fa8f..2ee3b68b6b 100644
+--- a/torch/csrc/jit/tensorexpr/llvm_codegen.h
+++ b/torch/csrc/jit/tensorexpr/llvm_codegen.h
+@@ -41,9 +41,9 @@ class TORCH_API LLVMCodeGen : public CodeGen {
+   // users can continue to call this kernel using `call` and `call_raw`.
+   void cleanup_memory();
+ 
+-  TORCH_API void call(const std::vector<CallArg>& args) override;
+-  TORCH_API void call_raw(const std::vector<void*>& args) override;
+-  TORCH_API void call_with_numel(void** args, int64_t numel) override;
+  void call(const std::vector<CallArg>& args) override;
+  void call_raw(const std::vector<void*>& args) override;
+  void call_with_numel(void** args, int64_t numel) override;
+ 
+   at::Tensor empty_strided(
+       c10::IntArrayRef size,
+diff --git a/torch/csrc/profiler/stubs/cuda.cpp b/torch/csrc/profiler/stubs/cuda.cpp
+index dec87576f3..3bb368a4df 100644
+--- a/torch/csrc/profiler/stubs/cuda.cpp
+++ b/torch/csrc/profiler/stubs/cuda.cpp
+@@ -1,6 +1,6 @@
+ #include <sstream>
+ 
+-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
+ 
+ #include <c10/cuda/CUDAGuard.h>
+ #include <c10/util/irange.h>
--- a/ports/libtorch/portfile.cmake
+++ b/ports/libtorch/portfile.cmake
@ -4,28 +4,67 @@ vcpkg_from_github(
    OUT_SOURCE_PATH SOURCE_PATH
    REPO pytorch/pytorch
    REF "v${VERSION}"
-    SHA512 afeb551904ebd9b5901ae623a98eadbb3045115247cedf8006a940742cfad04e5ce24cfaf363336a9ed88d7ce6a4ac53dbb6a5c690aef6efdf20477c3a22c7ca
+    SHA512 a8961d78ad785b13c959a0612563a60e0de17a7c8bb9822ddea9a24072796354d07e81c47b6cc8761b21a6448845b088cf80e1661d9e889b0ed5474d3dc76756
    HEAD_REF master
    PATCHES
-        missing_include.patch
-        pytorch-pr-85958.patch # https://github.com/pytorch/pytorch/pull/85958
-        fix-cmake.patch
-        fix-fbgemm-include.patch
-        fix-c10-glog.patch
-        use-flatbuffers2.patch # check with codegen-flatc-mobile_bytecode
-        fix-windows.patch # https://github.com/pytorch/pytorch/issues/87957
-        fix_werror.patch
+        cmake-fixes.patch
+        more-fixes.patch
+        fix-build.patch
+        clang-cl.patch
+        cuda-adjustments.patch
+        fix-api-export.patch
+        fix-onnx-case.patch
+        fxdiv.patch
+        protoc.patch
+        fix-sleef.patch
+        fix-glog.patch
+        fix-msvc-ICE.patch
 )
 file(REMOVE_RECURSE "${SOURCE_PATH}/caffe2/core/macros.h") # We must use generated header files

-# Editing ${SOURCE_PATH}/cmake/Dependencies.cmake makes HORRIBLE readability...
-file(COPY "${CMAKE_CURRENT_LIST_DIR}/vcpkg-dependencies.cmake" DESTINATION "${SOURCE_PATH}/cmake")
+vcpkg_from_github(
+    OUT_SOURCE_PATH src_kineto
+    REPO pytorch/kineto
+    REF 49e854d805d916b2031e337763928d2f8d2e1fbf
+    SHA512 ae63d48dc5b8ac30c38c2ace60f16834c7e9275fa342dc9f109d4fbc87b7bd674664f6413c36d0c1ab5a7da786030a4108d83daa4502b2f30239283ea3acdb16
+    HEAD_REF main
+    PATCHES
+      kineto.patch
+)
+file(COPY "${src_kineto}/" DESTINATION "${SOURCE_PATH}/third_party/kineto")
+
+vcpkg_from_github(
+    OUT_SOURCE_PATH src_cudnn
+    REPO NVIDIA/cudnn-frontend # new port ?
+    REF 12f35fa2be5994c1106367cac2fba21457b064f4
+    SHA512 a7e4bf58f82ca0b767df35da1b3588e2639ea2ef22ed0c47e989fb4cde5a28b0605b228b42fcaefbdf721bfbb91f2a9e7d41352ff522bd80b63db6d27e44ec20
+    HEAD_REF main
+)
+file(COPY "${src_cudnn}/" DESTINATION "${SOURCE_PATH}/third_party/cudnn_frontend")
+
+
+vcpkg_from_github(
+    OUT_SOURCE_PATH src_cutlass
+    REPO NVIDIA/cutlass # new port ?
+    REF 6f47420213f757831fae65c686aa471749fa8d60
+    SHA512 f3b3c43fbd7942f96407669405385c9a99274290e99f86cab5bb8657664bf1951e4da27f3069500a4825c427adeec883e05e81302b58390df3a3adb8c08e31ed
+    HEAD_REF main
+)
+file(COPY "${src_cutlass}/" DESTINATION "${SOURCE_PATH}/third_party/cutlass")
+
+file(REMOVE
+  "${SOURCE_PATH}/cmake/Modules/FindBLAS.cmake"
+  "${SOURCE_PATH}/cmake/Modules/FindLAPACK.cmake"
+  "${SOURCE_PATH}/cmake/Modules/FindCUDA.cmake"
+  "${SOURCE_PATH}/cmake/Modules/FindCUDAToolkit.cmake"
+  "${SOURCE_PATH}/cmake/Modules/Findpybind11.cmake"
+)

 find_program(FLATC NAMES flatc PATHS "${CURRENT_HOST_INSTALLED_DIR}/tools/flatbuffers" REQUIRED NO_DEFAULT_PATH NO_CMAKE_PATH)
 message(STATUS "Using flatc: ${FLATC}")

 vcpkg_execute_required_process(
-    COMMAND ${FLATC} --cpp --gen-object-api --gen-mutable mobile_bytecode.fbs
+    COMMAND ${FLATC} --cpp --no-prefix --scoped-enums --gen-mutable mobile_bytecode.fbs
    LOGNAME codegen-flatc-mobile_bytecode
    WORKING_DIRECTORY "${SOURCE_PATH}/torch/csrc/jit/serialization"
 )
@ -38,6 +77,7 @@ x_vcpkg_get_python_packages(
    PACKAGES typing-extensions pyyaml numpy
    OUT_PYTHON_VAR PYTHON3
 )
+#set(PYTHON3 "${CURRENT_HOST_INSTALLED_DIR}/tools/python3/python${VCPKG_HOST_EXECUTABLE_SUFFIX}")
 message(STATUS "Using Python3: ${PYTHON3}")

 vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS
@ -48,10 +88,11 @@ vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS
    fftw3   AT_FFTW_ENABLED
    fbgemm  USE_FBGEMM
    opencv  USE_OPENCV
-    tbb     USE_TBB
-    tbb     AT_PARALLEL_NATIVE_TBB
-    openmp  USE_OPENMP
-    openmp  AT_PARALLEL_OPENMP
+    # These are alternatives !
+    # tbb     USE_TBB
+    # tbb     AT_PARALLEL_NATIVE_TBB # AT_PARALLEL_ are alternatives
+    # openmp  USE_OPENMP
+    # openmp  AT_PARALLEL_OPENMP # AT_PARALLEL_ are alternatives
    leveldb USE_LEVELDB
    opencl  USE_OPENCL
    cuda    USE_CUDA
@ -61,20 +102,22 @@ vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS
    cuda    USE_NVRTC
    cuda    AT_CUDA_ENABLED
    cuda    AT_CUDNN_ENABLED
+    cuda    USE_MAGMA
    vulkan  USE_VULKAN
-    vulkan  USE_VULKAN_WRAPPER
-    vulkan  USE_VULKAN_SHADERC_RUNTIME
+    #vulkan  USE_VULKAN_SHADERC_RUNTIME
    vulkan  USE_VULKAN_RELAXED_PRECISION
+    rocm    USE_ROCM  # This is an alternative to cuda not a feature! (Not in vcpkg.json!) -> disabled
+    llvm    USE_LLVM
+    mpi     USE_MPI
    nnpack  USE_NNPACK  # todo: check use of `DISABLE_NNPACK_AND_FAMILY`
    nnpack  AT_NNPACK_ENABLED
    xnnpack USE_XNNPACK
-    xnnpack USE_SYSTEM_XNNPACK
    qnnpack USE_QNNPACK # todo: check use of `USE_PYTORCH_QNNPACK`
+#   No feature in vcpkg yet so disabled. -> Requires numpy build by vcpkg itself
+    python  BUILD_PYTHON
+    python  USE_NUMPY
 )

-if(CMAKE_CXX_COMPILER_ID MATCHES GNU)
-    list(APPEND FEATURE_OPTIONS -DUSE_NATIVE_ARCH=ON)
-endif()
 if("dist" IN_LIST FEATURES)
    if(VCPKG_TARGET_IS_LINUX OR VCPKG_TARGET_IS_OSX)
        list(APPEND FEATURE_OPTIONS -DUSE_TENSORPIPE=ON)
@ -85,20 +128,6 @@ if("dist" IN_LIST FEATURES)
    list(APPEND FEATURE_OPTIONS -DUSE_GLOO=${VCPKG_TARGET_IS_LINUX})
 endif()

-# BLAS: MKL, OpenBLAS, or Accelerate
-#   The feature will be disabled if "generic" BLAS is not found
-if(VCPKG_TARGET_IS_OSX OR VCPKG_TARGET_IS_IOS)
-    list(APPEND BLAS_OPTIONS -DBLAS=Accelerate -DUSE_BLAS=ON)
-elseif(VCPKG_TARGET_IS_WINDOWS)
-    list(APPEND BLAS_OPTIONS -DBLAS=OpenBLAS -DUSE_BLAS=ON)
-elseif(VCPKG_TARGET_IS_LINUX)
-    list(APPEND BLAS_OPTIONS -DBLAS=generic -DUSE_BLAS=ON)
-endif()
-
-if("tbb" IN_LIST FEATURES)
-    list(APPEND FEATURE_OPTIONS -DMKLDNN_CPU_RUNTIME=TBB)
-endif()
-
 if(VCPKG_TARGET_IS_ANDROID OR VCPKG_TARGET_IS_IOS)
    list(APPEND FEATURE_OPTIONS -DINTERN_BUILD_MOBILE=ON)
 else()
@ -115,51 +144,76 @@ vcpkg_cmake_configure(
        -DProtobuf_PROTOC_EXECUTABLE:FILEPATH=${PROTOC}
        -DCAFFE2_CUSTOM_PROTOC_EXECUTABLE:FILEPATH=${PROTOC}
        -DPYTHON_EXECUTABLE:FILEPATH=${PYTHON3}
-        -DPython3_EXECUTABLE:FILEPATH=${PYTHON3}
+        #-DPython3_EXECUTABLE:FILEPATH=${PYTHON3}
+        -DCAFFE2_STATIC_LINK_CUDA=ON
        -DCAFFE2_USE_MSVC_STATIC_RUNTIME=${USE_STATIC_RUNTIME}
        -DBUILD_CUSTOM_PROTOBUF=OFF
        -DUSE_LITE_PROTO=OFF
        -DBUILD_TEST=OFF
        -DATEN_NO_TEST=ON
        -DUSE_SYSTEM_LIBS=ON
-        -DBUILD_PYTHON=OFF
-        -DUSE_NUMPY=OFF
        -DUSE_METAL=OFF
        -DUSE_PYTORCH_METAL=OFF
        -DUSE_PYTORCH_METAL_EXPORT=OFF
        -DUSE_GFLAGS=ON
        -DUSE_GLOG=ON
        -DUSE_LMDB=ON
-        -DUSE_ROCKSDB=OFF
-        -DUSE_OBSERVERS=OFF 
+        -DUSE_ITT=OFF
+        -DUSE_ROCKSDB=ON
+        -DUSE_OBSERVERS=OFF
        -DUSE_PYTORCH_QNNPACK=OFF
        -DUSE_KINETO=OFF
        -DUSE_ROCM=OFF
-        -DUSE_DEPLOY=OFF
-        -DUSE_FFTW=OFF
        -DUSE_NUMA=OFF
-        -DUSE_MPI=${VCPKG_TARGET_IS_LINUX}
+        -DUSE_SYSTEM_ONNX=ON
+        -DUSE_SYSTEM_FP16=ON
+        -DUSE_SYSTEM_EIGEN_INSTALL=ON
+        -DUSE_SYSTEM_CPUINFO=ON
+        -DUSE_SYSTEM_PTHREADPOOL=ON
+        -DUSE_SYSTEM_PYBIND11=ON
+        -DUSE_SYSTEM_ZSTD=ON
+        -DUSE_SYSTEM_XNNPACK=ON
+        -DUSE_SYSTEM_GLOO=ON
+        -DUSE_SYSTEM_NCCL=ON
+        -DUSE_SYSTEM_LIBS=ON
+        -DUSE_SYSTEM_FXDIV=ON
+        -DUSE_SYSTEM_SLEEF=ON
        -DBUILD_JNI=${VCPKG_TARGET_IS_ANDROID}
        -DUSE_NNAPI=${VCPKG_TARGET_IS_ANDROID}
        ${BLAS_OPTIONS}
        # BLAS=MKL not supported in this port
        -DUSE_MKLDNN=OFF
        -DUSE_MKLDNN_CBLAS=OFF
-        -DCAFFE2_USE_MKL=OFF
-        -DCAFFE2_USE_MKLDNN=OFF
-        -DAT_MKL_ENABLED=OFF
+        #-DCAFFE2_USE_MKL=ON
+        #-DAT_MKL_ENABLED=ON
        -DAT_MKLDNN_ENABLED=OFF
+        -DUSE_OPENCL=ON
+        -DUSE_NUMPY=ON
+        -DUSE_KINETO=OFF #
+    OPTIONS_RELEASE
+      -DPYTHON_LIBRARY=${CURRENT_INSTALLED_DIR}/lib/python311.lib
+    OPTIONS_DEBUG
+      -DPYTHON_LIBRARY=${CURRENT_INSTALLED_DIR}/debug/lib/python311_d.lib
    MAYBE_UNUSED_VARIABLES
        USE_NUMA
        USE_SYSTEM_BIND11
-        USE_VULKAN_WRAPPER
        MKLDNN_CPU_RUNTIME
+        PYTHON_LIBRARY
 )
-vcpkg_cmake_build(TARGET __aten_op_header_gen LOGFILE_BASE build-header_gen) # explicit codegen is required
-vcpkg_cmake_build(TARGET torch_cpu  LOGFILE_BASE build-torch_cpu)
+
 vcpkg_cmake_install()
 vcpkg_copy_pdbs()

+vcpkg_cmake_config_fixup(PACKAGE_NAME caffe2 CONFIG_PATH "share/cmake/Caffe2" DO_NOT_DELETE_PARENT_CONFIG_PATH)
+vcpkg_cmake_config_fixup(PACKAGE_NAME torch CONFIG_PATH "share/cmake/Torch")
+vcpkg_replace_string("${CURRENT_PACKAGES_DIR}/share/torch/TorchConfig.cmake" "/../../../" "/../../")
+vcpkg_replace_string("${CURRENT_PACKAGES_DIR}/share/caffe2/Caffe2Config.cmake" "/../../../" "/../../")
+vcpkg_replace_string("${CURRENT_PACKAGES_DIR}/share/caffe2/Caffe2Config.cmake"
+  "set(Caffe2_MAIN_LIBS torch_library)"
+  "set(Caffe2_MAIN_LIBS torch_library)\nfind_dependency(Eigen3)")
+
+
+
 # Traverse the folder and remove "some" empty folders
 function(cleanup_once folder)
    if(NOT IS_DIRECTORY "${folder}")
@ -186,16 +240,37 @@ function(cleanup_repeat folder repeat)
        return()
    endif()
    while(repeat GREATER_EQUAL 1)
-        math(EXPR repeat "${repeat} - 1" OUTPUT_FORMAT DECIMAL)   
+        math(EXPR repeat "${repeat} - 1" OUTPUT_FORMAT DECIMAL)
        cleanup_once("${folder}")
    endwhile()
 endfunction()

+cleanup_repeat("${CURRENT_PACKAGES_DIR}/include" 5)
+cleanup_repeat("${CURRENT_PACKAGES_DIR}/lib/site-packages" 13)
+
 file(REMOVE_RECURSE
    "${CURRENT_PACKAGES_DIR}/debug/include"
    "${CURRENT_PACKAGES_DIR}/debug/share"
-    "${CURRENT_PACKAGES_DIR}/share/cmake/ATen"
 )
-cleanup_repeat("${CURRENT_PACKAGES_DIR}/include" 5)

 vcpkg_install_copyright(FILE_LIST "${SOURCE_PATH}/LICENSE")
+
+# Cannot build python bindings yet
+#if("python" IN_LIST FEATURES)
+#  set(ENV{USE_SYSTEM_LIBS} 1)
+#  vcpkg_replace_string("${SOURCE_PATH}/setup.py" "@TARGET_TRIPLET@" "${TARGET_TRIPLET}-rel")
+#  vcpkg_replace_string("${SOURCE_PATH}/tools/setup_helpers/env.py" "@TARGET_TRIPLET@" "${TARGET_TRIPLET}-rel")
+#  vcpkg_replace_string("${SOURCE_PATH}/torch/utils/cpp_extension.py" "@TARGET_TRIPLET@" "${TARGET_TRIPLET}-rel")
+#  vcpkg_python_build_and_install_wheel(SOURCE_PATH "${SOURCE_PATH}" OPTIONS -x)
+#endif()
+
+set(VCPKG_POLICY_DLLS_WITHOUT_EXPORTS enabled) # torch_global_deps.dll is empty.c and just for linking deps
+
+set(config "${CURRENT_PACKAGES_DIR}/share/torch/TorchConfig.cmake")
+file(READ "${config}" contents)
+string(REGEX REPLACE "set\\\(NVTOOLEXT_HOME[^)]+" "set(NVTOOLEXT_HOME \"\$ENV{CUDA_PATH}\"" contents "${contents}")
+#string(REGEX REPLACE "set\\\(NVTOOLEXT_HOME[^)]+" "set(NVTOOLEXT_HOME \"\${CMAKE_CURRENT_LIST_DIR}/../../tools/cuda/\"" contents "${contents}")
+string(REGEX REPLACE "\\\${NVTOOLEXT_HOME}/lib/x64/nvToolsExt64_1.lib" "" contents "${contents}")
+file(WRITE "${config}" "${contents}")
+
+vcpkg_replace_string("${CURRENT_PACKAGES_DIR}/include/torch/csrc/autograd/custom_function.h" "struct TORCH_API Function" "struct Function")
--- a/ports/libtorch/protoc.patch
+++ b/ports/libtorch/protoc.patch
@ -0,0 +1,13 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index fa6da328b..80bbfc6bb 100644
+--- a/CMakeLists.txt	
+++ b/CMakeLists.txt
+@@ -686,7 +686,7 @@ if(NOT CMAKE_BUILD_TYPE)
+ endif()
+ 
+ # The below means we are cross compiling for arm64 or x86_64 on MacOSX
+-if(NOT IOS AND CMAKE_SYSTEM_NAME STREQUAL "Darwin" AND CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64|arm64)$")
+if(0)
+   set(CROSS_COMPILING_MACOSX TRUE)
+   # We need to compile a universal protoc to not fail protobuf build
+   # We set CMAKE_TRY_COMPILE_TARGET_TYPE to STATIC_LIBRARY (vs executable) to succeed the cmake compiler check for cross-compiling
--- a/ports/libtorch/use-flatbuffers2.patch
+++ b/ports/libtorch/use-flatbuffers2.patch
@ -1,276 +0,0 @@
-diff --git a/torch/csrc/jit/mobile/flatbuffer_loader.cpp b/torch/csrc/jit/mobile/flatbuffer_loader.cpp
-index 0e126ff..4e6048e 100644
--- a/torch/csrc/jit/mobile/flatbuffer_loader.cpp
-+++ b/torch/csrc/jit/mobile/flatbuffer_loader.cpp
-@@ -123,29 +123,29 @@ FlatbufferLoader::FlatbufferLoader()
-     : mcu_(std::make_shared<mobile::CompilationUnit>()),
-       cu_(std::make_shared<CompilationUnit>()),
-       ivalue_parsers_{nullptr} {
-  registerIValueParser(mobile::serialization::IValueUnion::NONE, &parseBasic);
-  registerIValueParser(mobile::serialization::IValueUnion::Int, &parseBasic);
-  registerIValueParser(mobile::serialization::IValueUnion::Bool, &parseBasic);
-  registerIValueParser(mobile::serialization::IValueUnion::Double, &parseBasic);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_NONE, &parseBasic);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_Int, &parseBasic);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_Bool, &parseBasic);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_Double, &parseBasic);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::ComplexDouble, &parseBasic);
-+      mobile::serialization::IValueUnion::IValueUnion_ComplexDouble, &parseBasic);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::TensorMetadata, &parseTensor);
-  registerIValueParser(mobile::serialization::IValueUnion::String, &parseBasic);
-  registerIValueParser(mobile::serialization::IValueUnion::List, &parseList);
-+      mobile::serialization::IValueUnion::IValueUnion_TensorMetadata, &parseTensor);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_String, &parseBasic);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_List, &parseList);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::IntList, &parseIntList);
-+      mobile::serialization::IValueUnion::IValueUnion_IntList, &parseIntList);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::DoubleList, &parseDoubleList);
-+      mobile::serialization::IValueUnion::IValueUnion_DoubleList, &parseDoubleList);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::BoolList, &parseBoolList);
-  registerIValueParser(mobile::serialization::IValueUnion::Tuple, &parseTuple);
-  registerIValueParser(mobile::serialization::IValueUnion::Dict, &parseDict);
-+      mobile::serialization::IValueUnion::IValueUnion_BoolList, &parseBoolList);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_Tuple, &parseTuple);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_Dict, &parseDict);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::Object, &parseObject);
-  registerIValueParser(mobile::serialization::IValueUnion::Device, &parseBasic);
-+      mobile::serialization::IValueUnion::IValueUnion_Object, &parseObject);
-+  registerIValueParser(mobile::serialization::IValueUnion::IValueUnion_Device, &parseBasic);
-   registerIValueParser(
-      mobile::serialization::IValueUnion::EnumValue, &parseEnum);
-+      mobile::serialization::IValueUnion::IValueUnion_EnumValue, &parseEnum);
-   internal_registerTypeResolver(&resolveType);
- }
- 
-@@ -334,21 +334,21 @@ IValue parseBasic(
-     FlatbufferLoader&,
-     const mobile::serialization::IValue& ivalue) {
-   switch (ivalue.val_type()) {
-    case mobile::serialization::IValueUnion::NONE:
-+    case mobile::serialization::IValueUnion::IValueUnion_NONE:
-       return {};
-    case mobile::serialization::IValueUnion::Int:
-+    case mobile::serialization::IValueUnion::IValueUnion_Int:
-       return ivalue.val_as_Int()->int_val();
-    case mobile::serialization::IValueUnion::Bool:
-+    case mobile::serialization::IValueUnion::IValueUnion_Bool:
-       return ivalue.val_as_Bool()->bool_val();
-    case mobile::serialization::IValueUnion::Double:
-+    case mobile::serialization::IValueUnion::IValueUnion_Double:
-       return ivalue.val_as_Double()->double_val();
-    case mobile::serialization::IValueUnion::ComplexDouble: {
-+    case mobile::serialization::IValueUnion::IValueUnion_ComplexDouble: {
-       const auto* comp = ivalue.val_as_ComplexDouble();
-       return c10::complex<double>(comp->real(), comp->imag());
-     }
-    case mobile::serialization::IValueUnion::String:
-+    case mobile::serialization::IValueUnion::IValueUnion_String:
-       return ivalue.val_as_String()->data()->str();
-    case mobile::serialization::IValueUnion::Device: {
-+    case mobile::serialization::IValueUnion::IValueUnion_Device: {
-       return c10::Device(ivalue.val_as_Device()->str()->str());
-     }
-     default:
-@@ -504,7 +504,7 @@ ClassTypePtr FlatbufferLoader::getOrCreateClassTypeForObject(
-     TORCH_CHECK(object->type_index() < all_ivalues_.size());
-     all_types_[object->type_index()] = cls;
- 
-    if (obj_type->type() == mobile::serialization::TypeType::CLASS_WITH_FIELD) {
-+    if (obj_type->type() == mobile::serialization::TypeType::TypeType_CLASS_WITH_FIELD) {
-       for (uint32_t i = 0; i < object->attrs()->size(); i++) {
-         IValue val = getIValue(object->attrs()->Get(i));
-         // Need to use concrete object's field's type to set type of field.
-@@ -529,7 +529,7 @@ IValue parseObject(
-   auto cls = loader.getOrCreateClassTypeForObject(object);
-   Stack stack;
-   switch (obj_type->type()) {
-    case mobile::serialization::TypeType::CLASS_WITH_FIELD: {
-+    case mobile::serialization::TypeType::TypeType_CLASS_WITH_FIELD: {
-       auto obj = c10::ivalue::Object::create(
-           at::StrongTypePtr(loader.cu_, cls), object->attrs()->size());
-       for (uint32_t i = 0; i < object->attrs()->size(); i++) {
-@@ -538,7 +538,7 @@ IValue parseObject(
-       }
-       return obj;
-     }
-    case mobile::serialization::TypeType::CLASS_WITH_SETSTATE: {
-+    case mobile::serialization::TypeType::TypeType_CLASS_WITH_SETSTATE: {
-       IValue input = loader.getIValue(object->state());
-       mobile::Function* setstate = loader.getFunction(object->setstate_func());
-       auto obj =
-@@ -548,7 +548,7 @@ IValue parseObject(
-       setstate->run(stack);
-       return obj;
-     }
-    case mobile::serialization::TypeType::CUSTOM_CLASS: {
-+    case mobile::serialization::TypeType::TypeType_CUSTOM_CLASS: {
-       auto custom_class_type =
-           torch::jit::getCustomClass(cls->name()->qualifiedName());
-       IValue input = loader.getIValue(object->state());
-diff --git a/torch/csrc/jit/mobile/flatbuffer_loader.h b/torch/csrc/jit/mobile/flatbuffer_loader.h
-index 7e88326..dc95202 100644
--- a/torch/csrc/jit/mobile/flatbuffer_loader.h
-+++ b/torch/csrc/jit/mobile/flatbuffer_loader.h
-@@ -151,7 +151,7 @@ class TORCH_API FlatbufferLoader {
-   std::vector<IValue> all_ivalues_;
-   std::array<
-       IValueParser,
-      static_cast<uint8_t>(mobile::serialization::IValueUnion::MAX) + 1>
-+      static_cast<uint8_t>(mobile::serialization::IValueUnion::IValueUnion_MAX) + 1>
-       ivalue_parsers_;
-   TypeResolver type_resolver_ = nullptr;
-   mobile::serialization::Module* module_ = nullptr;
-diff --git a/torch/csrc/jit/serialization/flatbuffer_serializer.cpp b/torch/csrc/jit/serialization/flatbuffer_serializer.cpp
-index 066a78c..a9e19bb 100644
--- a/torch/csrc/jit/serialization/flatbuffer_serializer.cpp
-+++ b/torch/csrc/jit/serialization/flatbuffer_serializer.cpp
-@@ -336,7 +336,7 @@ flatbuffers::DetachedBuffer FlatbufferSerializer::serializeModule(
-   mcu_ = &module.compilation_unit();
- 
-   // first element is None.
-  insertIValue(CreateIValue(fbb, mobile::serialization::IValueUnion::NONE, 0));
-+  insertIValue(CreateIValue(fbb, mobile::serialization::IValueUnion::IValueUnion_NONE, 0));
- 
-   auto methods = module.get_methods();
-   std::vector<uint32_t> functions_index;
-@@ -459,7 +459,7 @@ flatbuffers::Offset<mobile::serialization::Dict> FlatbufferSerializer::dictToFB(
- flatbuffers::Offset<mobile::serialization::ObjectType> FlatbufferSerializer::
-     classTypeToFB(FlatBufferBuilder& fbb, ClassTypePtr class_ptr) {
-   mobile::serialization::TypeType typetype =
-      mobile::serialization::TypeType::UNSET;
-+      mobile::serialization::TypeType::TypeType_UNSET;
- 
-   flatbuffers::Offset<
-       flatbuffers::Vector<flatbuffers::Offset<flatbuffers::String>>>
-@@ -469,11 +469,11 @@ flatbuffers::Offset<mobile::serialization::ObjectType> FlatbufferSerializer::
-   const mobile::Function* setstate = mcu_->find_function(setstate_name);
-   const mobile::Function* getstate = mcu_->find_function(getstate_name);
-   if (setstate != nullptr && getstate != nullptr) {
-    typetype = mobile::serialization::TypeType::CLASS_WITH_SETSTATE;
-+    typetype = mobile::serialization::TypeType::TypeType_CLASS_WITH_SETSTATE;
-   } else if (
-       class_ptr->findMethod("__setstate__") &&
-       class_ptr->findMethod("__getstate__")) {
-    typetype = mobile::serialization::TypeType::CUSTOM_CLASS;
-+    typetype = mobile::serialization::TypeType::TypeType_CUSTOM_CLASS;
-   } else {
-     size_t num_attr = class_ptr->numAttributes();
-     std::vector<flatbuffers::Offset<flatbuffers::String>> names;
-@@ -482,7 +482,7 @@ flatbuffers::Offset<mobile::serialization::ObjectType> FlatbufferSerializer::
-       names.push_back(fbb.CreateSharedString(class_ptr->getAttributeName(i)));
-     }
-     names_offset = fbb.CreateVector(names);
-    typetype = mobile::serialization::TypeType::CLASS_WITH_FIELD;
-+    typetype = mobile::serialization::TypeType::TypeType_CLASS_WITH_FIELD;
-   }
- 
-   auto name_offset = fbb.CreateString(class_ptr->name()->qualifiedName());
-@@ -500,7 +500,7 @@ uint32_t FlatbufferSerializer::storeFunctionAndGetIndex(
- 
-   auto offset = CreateIValue(
-       fbb,
-      mobile::serialization::IValueUnion::Function,
-+      mobile::serialization::IValueUnion::IValueUnion_Function,
-       functionToFB(fbb, qn, function).Union());
- 
-   uint32_t index = insertIValue(offset);
-@@ -662,68 +662,68 @@ flatbuffers::Offset<mobile::serialization::IValue> FlatbufferSerializer::
-     iValueToFB(flatbuffers::FlatBufferBuilder& fbb, const IValue& ivalue) {
-   using mobile::serialization::IValueUnion;
- 
-  IValueUnion ivalue_type = IValueUnion::NONE;
-+  IValueUnion ivalue_type = IValueUnion::IValueUnion_NONE;
-   flatbuffers::Offset<void> offset = 0;
- 
-   if (ivalue.isTensor()) {
-    ivalue_type = IValueUnion::TensorMetadata;
-+    ivalue_type = IValueUnion::IValueUnion_TensorMetadata;
-     offset = tensorToFB(fbb, ivalue).Union();
-   } else if (ivalue.isTuple()) {
-    ivalue_type = IValueUnion::Tuple;
-+    ivalue_type = IValueUnion::IValueUnion_Tuple;
-     offset = tupleToFB(fbb, ivalue).Union();
-   } else if (ivalue.isDouble()) {
-    ivalue_type = IValueUnion::Double;
-+    ivalue_type = IValueUnion::IValueUnion_Double;
-     offset = fbb.CreateStruct(mobile::serialization::Double(ivalue.toDouble()))
-                  .Union();
-   } else if (ivalue.isComplexDouble()) {
-     auto comp = ivalue.toComplexDouble();
-    ivalue_type = IValueUnion::ComplexDouble;
-+    ivalue_type = IValueUnion::IValueUnion_ComplexDouble;
-     offset = fbb.CreateStruct(mobile::serialization::ComplexDouble(
-                                   comp.real(), comp.imag()))
-                  .Union();
-   } else if (ivalue.isInt()) {
-    ivalue_type = IValueUnion::Int;
-+    ivalue_type = IValueUnion::IValueUnion_Int;
-     offset =
-         fbb.CreateStruct(mobile::serialization::Int(ivalue.toInt())).Union();
-   } else if (ivalue.isBool()) {
-    ivalue_type = IValueUnion::Bool;
-+    ivalue_type = IValueUnion::IValueUnion_Bool;
-     offset =
-         fbb.CreateStruct(mobile::serialization::Bool(ivalue.toBool())).Union();
-   } else if (ivalue.isString()) {
-    ivalue_type = IValueUnion::String;
-+    ivalue_type = IValueUnion::IValueUnion_String;
-     offset = mobile::serialization::CreateString(
-                  fbb, fbb.CreateSharedString(ivalue.toString()->string()))
-                  .Union();
-   } else if (ivalue.isGenericDict()) {
-    ivalue_type = IValueUnion::Dict;
-+    ivalue_type = IValueUnion::IValueUnion_Dict;
-     offset = dictToFB(fbb, ivalue).Union();
-   } else if (ivalue.isNone()) {
-    ivalue_type = IValueUnion::NONE;
-+    ivalue_type = IValueUnion::IValueUnion_NONE;
-     offset = 0;
-   } else if (ivalue.isIntList()) {
-    ivalue_type = IValueUnion::IntList;
-+    ivalue_type = IValueUnion::IValueUnion_IntList;
-     offset = mobile::serialization::CreateIntList(
-                  fbb, fbb.CreateVector(ivalue.toIntVector()))
-                  .Union();
-   } else if (ivalue.isDoubleList()) {
-    ivalue_type = IValueUnion::DoubleList;
-+    ivalue_type = IValueUnion::IValueUnion_DoubleList;
-     offset = mobile::serialization::CreateDoubleList(
-                  fbb, fbb.CreateVector(ivalue.toDoubleVector()))
-                  .Union();
-   } else if (ivalue.isBoolList()) {
-    ivalue_type = IValueUnion::BoolList;
-+    ivalue_type = IValueUnion::IValueUnion_BoolList;
-     auto boollist = ivalue.toBoolList();
-     std::vector<uint8_t> bool_vec(boollist.begin(), boollist.end());
-     offset =
-         mobile::serialization::CreateBoolListDirect(fbb, &bool_vec).Union();
-   } else if (ivalue.isList()) {
-    ivalue_type = IValueUnion::List;
-+    ivalue_type = IValueUnion::IValueUnion_List;
-     offset = listToFB(fbb, ivalue).Union();
-   } else if (ivalue.isObject()) {
-    ivalue_type = IValueUnion::Object;
-+    ivalue_type = IValueUnion::IValueUnion_Object;
-     offset = objectToFB(fbb, ivalue).Union();
-   } else if (ivalue.isDevice()) {
-    ivalue_type = IValueUnion::Device;
-+    ivalue_type = IValueUnion::IValueUnion_Device;
-     offset = mobile::serialization::CreateDevice(
-                  fbb, fbb.CreateSharedString(ivalue.toDevice().str()))
-                  .Union();
-@@ -732,7 +732,7 @@ flatbuffers::Offset<mobile::serialization::IValue> FlatbufferSerializer::
-     const auto& qualified_class_name =
-         enum_holder->type()->qualifiedClassName();
-     uint32_t ival_pos = storeIValueAndGetIndex(fbb, enum_holder->value());
-    ivalue_type = IValueUnion::EnumValue;
-+    ivalue_type = IValueUnion::IValueUnion_EnumValue;
-     offset = mobile::serialization::CreateEnumValue(
-                  fbb,
-                  fbb.CreateSharedString(qualified_class_name.qualifiedName()),
--- a/ports/libtorch/vcpkg-dependencies.cmake
+++ b/ports/libtorch/vcpkg-dependencies.cmake
@ -1,263 +0,0 @@
-# caffe2/core/macros.h.in
-set(CAFFE2_USE_GOOGLE_GLOG 1)
-set(CAFFE2_USE_LITE_PROTO 0)
-# set(CAFFE2_FORCE_FALLBACK_CUDA_MPI 0)
-
-# aten/src/ATen/Config.h.in
-set(AT_POCKETFFT_ENABLED 0)
-set(AT_MKL_ENABLED 0)
-set(AT_FFTW_ENABLED 0)
-set(AT_PARALLEL_NATIVE 1)
-if(USE_OPENMP OR USE_TBB)
-  set(AT_PARALLEL_NATIVE 0)
-endif()
-
-find_program(Python3_EXECUTABLE NAMES python3 python REQUIRED)
-find_program(PYTHON_EXECUTABLE NAMES python3 python REQUIRED)
-
-find_package(Python3 REQUIRED COMPONENTS Interpreter)
-if(BUILD_PYTHON)
-    find_package(Python3 REQUIRED COMPONENTS Development NumPy)
-    list(APPEND Caffe2_DEPENDENCY_LIBS Python3::Python Python3::NumPy)
-    find_package(pybind11 CONFIG REQUIRED)
-    list(APPEND Caffe2_DEPENDENCY_LIBS pybind11::pybind11)
-endif()
-
-#Caffe2 depends on FP16 library for half-precision conversions
-find_path(FP16_INCLUDE_DIRS "fp16.h" REQUIRED)
-# PSIMD is required by FP16
-find_path(PSIMD_INCLUDE_DIRS "psimd.h" REQUIRED)
-if(USE_NNPACK OR USE_QNNPACK OR USE_PYTORCH_QNNPACK OR USE_XNNPACK)
-    find_path(FXDIV_INCLUDE_DIRS "fxdiv.h" REQUIRED)
-endif()
-
-find_library(FOXI_LOADER_LIBPATH NAMES foxi_loader REQUIRED)
-list(APPEND Caffe2_DEPENDENCY_LIBS ${FOXI_LOADER_LIBPATH})
-
-find_package(Threads REQUIRED) # Threads::Threads
-find_package(gemmlowp CONFIG REQUIRED) # gemmlowp::gemmlowp
-find_package(gflags CONFIG REQUIRED) # gflags::gflags
-find_package(glog CONFIG REQUIRED) # glog::glog
-find_package(cpuinfo CONFIG REQUIRED) # cpuinfo::clog cpuinfo::cpuinfo
-list(APPEND Caffe2_DEPENDENCY_LIBS gemmlowp::gemmlowp gflags::gflags glog::glog cpuinfo::clog cpuinfo::cpuinfo)
-
-# Only add a dependency on pthreadpool if we are on a mobile build or are building any of the libraries in the {Q/X}NNPACK family.
-if(INTERN_BUILD_MOBILE OR USE_NNPACK OR USE_QNNPACK OR USE_PYTORCH_QNNPACK OR USE_XNNPACK)
-    find_package(unofficial-pthreadpool CONFIG REQUIRED) # unofficial::pthreadpool
-    list(APPEND Caffe2_DEPENDENCY_LIBS unofficial::pthreadpool)
-    set(USE_PTHREADPOOL 1)
-    set(USE_INTERNAL_PTHREADPOOL_IMPL 0)
-    add_compile_definitions(USE_PTHREADPOOL)
-endif()
-
-link_directories(
-  $<$<CONFIG:Debug>:${_VCPKG_INSTALLED_DIR}/${VCPKG_TARGET_TRIPLET}/debug/lib>
-  $<$<CONFIG:Release>:${_VCPKG_INSTALLED_DIR}/${VCPKG_TARGET_TRIPLET}/lib>
-)
-
-find_package(fmt CONFIG REQUIRED) # fmt::fmt-header-only
-list(APPEND Caffe2_DEPENDENCY_LIBS fmt::fmt-header-only)
-
-find_package(Eigen3 CONFIG REQUIRED) # Eigen3::Eigen
-include_directories(SYSTEM ${EIGEN3_INCLUDE_DIR})
-list(APPEND Caffe2_DEPENDENCY_LIBS Eigen3::Eigen)
-set(CAFFE2_USE_EIGEN_FOR_BLAS 1) # see caff2/core/macros.h.in
-
-if(BLAS STREQUAL "Accelerate")
-  set(WITH_BLAS "accelerate")
-  find_package(BLAS REQUIRED) # cmake/Modules/FindBLAS.cmake
-  find_package(LAPACK REQUIRED) # cmake/Modules/FindLAPACK.cmake
-  set(USE_LAPACK 1)
-  list(APPEND Caffe2_PRIVATE_DEPENDENCY_LIBS ${LAPACK_LIBRARIES})
-
-elseif(BLAS STREQUAL "Open")
-  set(WITH_BLAS "open")
-  find_package(BLAS)
-  if(BLAS_FOUND)
-    set(USE_BLAS 1)
-    list(APPEND Caffe2_PRIVATE_DEPENDENCY_LIBS ${BLAS_LIBRARIES})
-  else()
-    set(USE_BLAS 0) # if we couldn't find the BLAS, disable the feature
-  endif()
-  set(USE_LAPACK 0)
-
-elseif(BLAS STREQUAL "MKL")
-  if(USE_TBB)
-    set(MKL_THREADING "TBB")
-  else()
-    set(MKL_THREADING "SEQ")
-  endif()
-  find_package(MKL REQUIRED)
-  include(${CMAKE_CURRENT_LIST_DIR}/public/mkl.cmake)
-  include_directories(AFTER SYSTEM ${MKL_INCLUDE_DIR})
-  list(APPEND Caffe2_PUBLIC_DEPENDENCY_LIBS caffe2::mkl)
-  set(WITH_BLAS "mkl")
-  find_package(BLAS REQUIRED) # cmake/Modules/FindBLAS.cmake
-
-else()
-  set(USE_BLAS 0)
-  set(WITH_BLAS "generic")
-  find_package(BLAS)
-  if(BLAS_FOUND)
-    set(USE_BLAS 1)
-    list(APPEND Caffe2_PRIVATE_DEPENDENCY_LIBS ${BLAS_LIBRARIES})
-  endif()
-
-endif()
-
-if(USE_MKLDNN)
-  find_package(MKLDNN REQUIRED) # BLAS::BLAS
-  include(cmake/public/mkldnn.cmake)
-  include_directories(AFTER SYSTEM ${MKLDNN_INCLUDE_DIR})
-  list(APPEND Caffe2_PUBLIC_DEPENDENCY_LIBS caffe2::mkldnn)
-endif()
-
-if(USE_TBB)
-  find_package(TBB CONFIG REQUIRED) # TBB::tbb
-  list(APPEND Caffe2_DEPENDENCY_LIBS TBB::tbb)
-endif()
-
-if(USE_NNPACK)
-  find_library(NNPACK_LIB NAMES nnpack REQUIRED)
-  list(APPEND Caffe2_DEPENDENCY_LIBS ${NNPACK_LIB})
-  string(APPEND CMAKE_CXX_FLAGS " -DUSE_NNPACK")
-endif()
-
-if(USE_FBGEMM)
-  find_package(asmjit CONFIG REQUIRED) # asmjit::asmjit (required by fbgemm)
-  find_package(fbgemmLibrary CONFIG REQUIRED) # fbgemm
-  list(APPEND Caffe2_DEPENDENCY_LIBS asmjit::asmjit fbgemm)
-  if(USE_CUDA)
-    # todo: fbgemm-gpu
-  endif()
-endif()
-
-if(USE_LMDB)
-  find_package(LMDB) # from cmake/Modules/FindLMDB.cmake
-  if(LMDB_FOUND)
-    list(APPEND Caffe2_DEPENDENCY_LIBS ${LMDB_LIBRARIES})
-  else()
-    find_package(lmdb CONFIG REQUIRED) # lmdb
-    list(APPEND Caffe2_DEPENDENCY_LIBS lmdb)
-  endif()
-endif()
-
-if(USE_LEVELDB)
-  find_package(Snappy CONFIG REQUIRED) # Snappy::snappy
-  list(APPEND Caffe2_DEPENDENCY_LIBS Snappy::snappy)
-  find_package(LevelDB) # from cmake/Modules/FindLevelDB.cmake
-  if(LevelDB_FOUND)
-    list(APPEND Caffe2_DEPENDENCY_LIBS ${LevelDB_LIBRARIES})
-  else()
-    find_package(leveldb CONFIG REQUIRED) # leveldb::leveldb
-    list(APPEND Caffe2_DEPENDENCY_LIBS leveldb::leveldb)
-  endif()
-endif()
- 
-if(USE_QNNPACK)
-  find_library(QNNPACK_LIB NAME qnnpack REQUIRED)
-  list(APPEND Caffe2_DEPENDENCY_LIBS ${QNNPACK_LIB})
-endif()
-
-if(USE_XNNPACK)
-  find_library(XNNPACK_LIBRARY NAMES xnnpack XNNPACK REQUIRED) # xnnpack
-  list(APPEND Caffe2_DEPENDENCY_LIBS ${XNNPACK_LIBRARY})
-endif()
- 
-if(USE_ZSTD)
-  find_package(zstd CONFIG REQUIRED) # zstd::libzstd_static
-  list(APPEND Caffe2_DEPENDENCY_LIBS $<IF:$<TARGET_EXISTS:zstd::libzstd_shared>,zstd::libzstd_shared,zstd::libzstd_static>)
-endif()
-
-if(USE_SYSTEM_ONNX)
-  find_package(ONNX CONFIG REQUIRED) # onnx onnx_proto onnxifi_loader
-  find_package(ONNXOptimizer CONFIG REQUIRED) # onnx_optimizer
-  list(APPEND Caffe2_DEPENDENCY_LIBS onnx onnx_proto onnx_optimizer)
-  if(ONNX_ML)
-    add_compile_definitions(ONNX_ML=1)
-  endif()
-endif()
-
-if(USE_CUDA)
-  set(CMAKE_CUDA_STANDARD 14)
-  find_package(CUDA  10.1 REQUIRED) # https://cmake.org/cmake/help/latest/module/FindCUDA.html
-  find_package(CUDNN 8.0  REQUIRED) # CuDNN::CuDNN
-  cuda_select_nvcc_arch_flags(ARCH_FLAGS 6.0 6.2 7.5)
-  list(APPEND CUDA_NVCC_FLAGS
-    ${ARCH_FLAGS}                # check TORCH_NVCC_FLAGS in this project
-    -D__CUDA_NO_HALF_OPERATORS__ # see https://github.com/torch/cutorch/issues/797
-  )
-  set(CAFFE2_USE_CUDA 1)
-  set(CAFFE2_USE_CUDNN 1)
-  set(CAFFE2_USE_NVRTC 0)
-  set(CAFFE2_USE_TENSORRT 0)
-  include(cmake/public/cuda.cmake)
-  list(APPEND Caffe2_DEPENDENCY_LIBS
-    CuDNN::CuDNN
-  )
-  list(APPEND Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS
-    caffe2::cufft caffe2::curand caffe2::cublas
-    caffe2::cudnn-public
-    # caffe2::cuda caffe2::nvrtc # for CAFFE2_USE_NVRTC
-    # caffe2::tensorrt # for CAFFE2_USE_TENSORRT
-  )
-endif()
-
-if(USE_NUMA) # Linux package 'libnuma-dev'
-  find_package(Numa REQUIRED)
-  include_directories(SYSTEM ${Numa_INCLUDE_DIR})
-  list(APPEND Caffe2_DEPENDENCY_LIBS ${Numa_LIBRARIES})
-endif()
-
-if(USE_GLOO)
-  find_package(Gloo CONFIG REQUIRED)
-  list(APPEND Caffe2_DEPENDENCY_LIBS gloo)
-endif()
-
-if(USE_VULKAN)
-  find_package(Vulkan REQUIRED)
-  list(APPEND Caffe2_DEPENDENCY_LIBS Vulkan::Vulkan)
-endif()
-
-if(USE_TENSORPIPE)
-  find_package(libuv CONFIG REQUIRED)
-  find_package(Tensorpipe CONFIG REQUIRED) # tensorpipe
-  list(APPEND Caffe2_DEPENDENCY_LIBS uv_a tensorpipe)
-endif()
-
-if(USE_MPI)
-  find_package(MPI REQUIRED) # APT package: libopenmpi-dev
-  if(NOT MPI_CXX_FOUND)
-    message(FATAL_ERROR "Failed to find MPI_CXX")
-  endif()
-  include_directories(SYSTEM ${MPI_CXX_INCLUDE_PATH})
-  list(APPEND Caffe2_DEPENDENCY_LIBS ${MPI_CXX_LIBRARIES})
-  list(APPEND CMAKE_EXE_LINKER_FLAGS ${MPI_CXX_LINK_FLAGS})
-
-  find_program(OMPI_INFO NAMES ompi_info HINTS ${MPI_CXX_LIBRARIES}/../bin)
-  if(OMPI_INFO)
-    execute_process(COMMAND ${OMPI_INFO} OUTPUT_VARIABLE _output)
-    if(_output MATCHES "smcuda")
-      message(STATUS "Found OpenMPI with CUDA support built.")
-    else()
-      message(WARNING "OpenMPI found, but it is not built with CUDA support.")
-      set(CAFFE2_FORCE_FALLBACK_CUDA_MPI 1)
-    endif()
-  endif()
-endif()
-
-if(USE_OPENCV)
-  find_package(OpenCV CONFIG REQUIRED COMPONENTS core highgui imgproc imgcodecs videoio video)
-  include_directories(SYSTEM ${OpenCV_INCLUDE_DIRS})
-  list(APPEND Caffe2_DEPENDENCY_LIBS ${OpenCV_LIBS})
-  if(MSVC AND USE_CUDA)
-    list(APPEND Caffe2_CUDA_DEPENDENCY_LIBS ${OpenCV_LIBS})
-  endif()
-endif()
-
-if(USE_OPENCL)
-  find_package(OpenCL REQUIRED)
-  include_directories(SYSTEM ${OpenCL_INCLUDE_DIRS})
-  include_directories(${CMAKE_CURRENT_LIST_DIR}/../caffe2/contrib/opencl)
-  list(APPEND Caffe2_DEPENDENCY_LIBS OpenCL::OpenCL)
-endif()
--- a/ports/libtorch/vcpkg.json
+++ b/ports/libtorch/vcpkg.json
@ -1,12 +1,12 @@
 {
  "name": "libtorch",
-  "version": "1.12.1",
-  "port-version": 5,
+  "version": "2.1.2",
  "description": "Tensors and Dynamic neural networks in Python with strong GPU acceleration",
  "homepage": "https://pytorch.org/",
  "license": null,
  "supports": "(windows & !static) | osx | linux",
  "dependencies": [
+    "blas",
    "cpuinfo",
    "eigen3",
    "flatbuffers",
@ -16,26 +16,24 @@
    },
    "fmt",
    "foxi",
+    "foxi",
    "fp16",
    "gemmlowp",
    "gflags",
    "glog",
+    "lapack",
    "lmdb",
-    {
-      "name": "mpi",
-      "platform": "linux"
-    },
+    "mimalloc",
    "onnx",
    "onnx-optimizer",
-    {
-      "name": "openblas",
-      "platform": "windows"
-    },
+    "opencl",
+    "pocketfft",
    "protobuf",
    {
      "name": "protobuf",
      "host": true
    },
+    "pthreadpool",
    "sleef",
    {
      "name": "vcpkg-cmake",
@ -51,9 +49,24 @@
    }
  ],
  "default-features": [
-    "xnnpack"
+    {
+      "name": "cuda",
+      "platform": "windows & !x86"
+    },
+    "fftw3",
+    "opencv",
+    "xnnpack",
+    "zstd"
  ],
  "features": {
+    "cuda": {
+      "description": "Build with CUDA GPU backend",
+      "dependencies": [
+        "cuda",
+        "cudnn",
+        "magma"
+      ]
+    },
    "dist": {
      "description": "Use distributed with MPI, Gloo, libuv, TensorPipe",
      "dependencies": [
@ -62,12 +75,16 @@
          "platform": "linux"
        },
        {
-          "name": "libuv",
-          "platform": "windows | osx"
+          "name": "libtorch",
+          "default-features": false,
+          "features": [
+            "mpi"
+          ],
+          "platform": "linux"
        },
        {
-          "name": "openmpi",
-          "platform": "linux | osx"
+          "name": "libuv",
+          "platform": "windows | osx"
        },
        {
          "name": "tensorpipe",
@ -75,6 +92,12 @@
        }
      ]
    },
+    "fftw3": {
+      "description": "Build with fftw3",
+      "dependencies": [
+        "fftw3"
+      ]
+    },
    "leveldb": {
      "description": "Build with LevelDB",
      "dependencies": [
@ -82,6 +105,18 @@
        "snappy"
      ]
    },
+    "llvm": {
+      "description": "Build with LLVM",
+      "dependencies": [
+        "llvm"
+      ]
+    },
+    "mpi": {
+      "description": "Build with MPI",
+      "dependencies": [
+        "mpi"
+      ]
+    },
    "nnpack": {
      "description": "Build with NNPack",
      "supports": "linux | osx",
@ -95,22 +130,21 @@
        "opencv"
      ]
    },
-    "tbb": {
-      "description": "Build with Intel TBB",
-      "dependencies": [
-        "tbb"
-      ]
-    },
    "vulkan": {
      "description": "Build with Vulkan GPU backend",
-      "supports": "!windows",
      "dependencies": [
-        "vulkan"
+        "vulkan",
+        "vulkan-loader",
+        "vulkan-memory-allocator"
      ]
    },
    "xnnpack": {
      "description": "Build with XNNPack",
      "dependencies": [
+        {
+          "name": "fxdiv",
+          "platform": "!windows"
+        },
        "xnnpack"
      ]
    },
--- a/ports/xnnpack/fix-build.patch
+++ b/ports/xnnpack/fix-build.patch
@ -0,0 +1,94 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 95df99eae0..7e969cfbb8 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -704,7 +704,7 @@ IF(MSVC)
+     TARGET_COMPILE_DEFINITIONS(operator-utils PRIVATE "restrict=")
+     TARGET_COMPILE_DEFINITIONS(XNNPACK PRIVATE "restrict=")
+   ENDIF()
+-
+  if(0)
+   TARGET_COMPILE_OPTIONS(microkernels-all PRIVATE "$<$<NOT:$<CONFIG:Debug>>:/O2>")
+   TARGET_COMPILE_OPTIONS(microkernels-prod PRIVATE "$<$<NOT:$<CONFIG:Debug>>:/O2>")
+   TARGET_COMPILE_OPTIONS(hardware-config PRIVATE "$<$<NOT:$<CONFIG:Debug>>:/O2>")
+@@ -723,6 +723,7 @@ IF(MSVC)
+     TARGET_COMPILE_OPTIONS(operators PRIVATE "$<$<NOT:$<CONFIG:Debug>>:/O1>")
+     TARGET_COMPILE_OPTIONS(XNNPACK PRIVATE "$<$<NOT:$<CONFIG:Debug>>:/O1>")
+   ENDIF()
+  endif()
+ ELSE()
+   TARGET_COMPILE_OPTIONS(microkernels-all PRIVATE "$<$<NOT:$<CONFIG:Debug>>:-O2>")
+   TARGET_COMPILE_OPTIONS(microkernels-prod PRIVATE "$<$<NOT:$<CONFIG:Debug>>:-O2>")
+@@ -797,8 +798,8 @@ IF(NOT TARGET clog)
+     # We build static version of clog but a dynamic library may indirectly depend on it
+     SET_PROPERTY(TARGET clog PROPERTY POSITION_INDEPENDENT_CODE ON)
+   ELSE()
+-    ADD_LIBRARY(clog STATIC IMPORTED)
+-    FIND_LIBRARY(CLOG_LIBRARY clog)
+    ADD_LIBRARY(clog UNKNOWN IMPORTED)
+    FIND_LIBRARY(CLOG_LIBRARY NAMES clog REQUIRED)
+     IF(NOT CLOG_LIBRARY)
+       MESSAGE(FATAL_ERROR "Cannot find clog")
+     ENDIF()
+@@ -817,8 +818,8 @@ IF(NOT TARGET cpuinfo)
+       "${CPUINFO_SOURCE_DIR}"
+       "${CMAKE_BINARY_DIR}/cpuinfo")
+   ELSE()
+-    ADD_LIBRARY(cpuinfo SHARED IMPORTED)
+-    FIND_LIBRARY(CPUINFO_LIBRARY cpuinfo)
+    ADD_LIBRARY(cpuinfo UNKNOWN IMPORTED)
+    FIND_LIBRARY(CPUINFO_LIBRARY NAMES cpuinfo REQUIRED)
+     IF(NOT CPUINFO_LIBRARY)
+       MESSAGE(FATAL_ERROR "Cannot find cpuinfo")
+     ENDIF()
+@@ -841,13 +842,15 @@ IF(NOT TARGET pthreadpool)
+       "${PTHREADPOOL_SOURCE_DIR}"
+       "${CMAKE_BINARY_DIR}/pthreadpool")
+   ELSE()
+-    ADD_LIBRARY(pthreadpool SHARED IMPORTED)
+-    FIND_LIBRARY(PTHREADPOOL_LIBRARY pthreadpool)
+    ADD_LIBRARY(pthreadpool UNKNOWN IMPORTED)
+    FIND_LIBRARY(PTHREADPOOL_LIBRARY NAMES pthreadpool REQUIRED)
+    find_path(PTHREADPOOL_INCLUDE NAMES pthreadpool.h REQUIRED PATH_SUFFIXES include)
+    target_include_directories(pthreadpool INTERFACE "${PTHREADPOOL_INCLUDE}")
+     IF(NOT PTHREADPOOL_LIBRARY)
+       MESSAGE(FATAL_ERROR "Cannot find pthreadpool")
+     ENDIF()
+     SET_PROPERTY(TARGET pthreadpool PROPERTY IMPORTED_LOCATION "${PTHREADPOOL_LIBRARY}")
+-    SET_PROPERTY(TARGET pthreadpool PROPERTY IMPORTED_IMPLIB "${PTHREADPOOL_LIBRARY}")
+    #SET_PROPERTY(TARGET pthreadpool PROPERTY IMPORTED_IMPLIB "${PTHREADPOOL_LIBRARY}")
+   ENDIF()
+ ENDIF()
+ TARGET_LINK_LIBRARIES(microkernels-all PRIVATE pthreadpool)
+@@ -881,12 +884,12 @@ IF(NOT TARGET fxdiv)
+       "${FXDIV_SOURCE_DIR}"
+       "${CMAKE_BINARY_DIR}/FXdiv")
+   ELSE()
+-    FIND_FILE(FXDIV_HDR fxdiv.h PATH_SUFFIXES include)
+    FIND_PATH(FXDIV_HDR fxdiv.h PATH_SUFFIXES include)
+     IF(NOT FXDIV_HDR)
+       MESSAGE(FATAL_ERROR "Cannot find fxdiv")
+     ENDIF()
+-    ADD_LIBRARY(fxdiv STATIC "${FXDIV_HDR}")
+-    SET_PROPERTY(TARGET fxdiv PROPERTY LINKER_LANGUAGE C)
+    ADD_LIBRARY(fxdiv INTERFACE IMPORTED)
+    target_include_directories(fxdiv INTERFACE "${FXDIV_HDR}")
+   ENDIF()
+ ENDIF()
+ TARGET_LINK_LIBRARIES(microkernels-all PRIVATE fxdiv)
+@@ -905,12 +908,12 @@ IF(NOT TARGET fp16)
+       "${FP16_SOURCE_DIR}"
+       "${CMAKE_BINARY_DIR}/FP16")
+   ELSE()
+-    FIND_FILE(FP16_HDR fp16.h PATH_SUFFIXES include)
+    FIND_PATH(FP16_HDR fp16.h PATH_SUFFIXES include REQUIRED)
+     IF(NOT FP16_HDR)
+       MESSAGE(FATAL_ERROR "Cannot find fp16")
+     ENDIF()
+-    ADD_LIBRARY(fp16 STATIC "${FP16_HDR}")
+-    SET_PROPERTY(TARGET fp16 PROPERTY LINKER_LANGUAGE C)
+    ADD_LIBRARY(fp16 INTERFACE IMPORTED)
+    target_include_directories(fp16 INTERFACE "${FP16_HDR}")
+   ENDIF()
+ ENDIF()
+ TARGET_LINK_LIBRARIES(microkernels-all PRIVATE fp16)
--- a/ports/xnnpack/portfile.cmake
+++ b/ports/xnnpack/portfile.cmake
@ -5,12 +5,11 @@ endif()
 vcpkg_from_github(
    OUT_SOURCE_PATH SOURCE_PATH
    REPO google/XNNPACK
-    REF ae108ef49aa5623b896fc93d4298c49d1750d9ba # 2022-02-17
-    SHA512 597354c8c5b786ba24a8c9759409fee7b090ec497547919da4c5786a571b5ef237d1da204b7fd553217089792acb5e40ea1e6d26e3a638aa32df1e63c14d1c91
+    REF 51a987591a6fc9f0fc0707077f53d763ac132cbf # 2022-12-22
+    SHA512 1742ad4de3aa1a67cd5a3cbf15d82afc2cc8569a2abac92542f748a440ef37fef4e9d9c302b7b2813e3176934b637c761c1c0956d20ab1ba28b280d9b121d91b
    HEAD_REF master
    PATCHES
-        use-packages.patch
-        use-c-cpp-11.patch
+        fix-build.patch
 )

 vcpkg_cmake_configure(
@ -25,7 +24,6 @@ vcpkg_cmake_configure(
 )
 vcpkg_cmake_install()
 vcpkg_copy_pdbs()
-vcpkg_cmake_config_fixup()

 file(INSTALL "${SOURCE_PATH}/LICENSE" DESTINATION "${CURRENT_PACKAGES_DIR}/share/${PORT}" RENAME copyright)
 file(REMOVE_RECURSE "${CURRENT_PACKAGES_DIR}/debug/include"
--- a/ports/xnnpack/use-c-cpp-11.patch
+++ b/ports/xnnpack/use-c-cpp-11.patch
@ -1,14 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index c924b45..72ffd64 100755
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -12,6 +12,9 @@ CMAKE_MINIMUM_REQUIRED(VERSION 3.12 FATAL_ERROR)
- # ---[ Project and semantic versioning.
- PROJECT(XNNPACK C CXX ASM)
- 
-+set(CMAKE_CXX_STANDARD 11)
-+set(CMAKE_C_STANDARD 11)
-+
- # ---[ Options.
- SET(XNNPACK_LIBRARY_TYPE "default" CACHE STRING "Type of library (shared, static, or default) to build")
- SET_PROPERTY(CACHE XNNPACK_LIBRARY_TYPE PROPERTY STRINGS default static shared)
--- a/ports/xnnpack/use-packages.patch
+++ b/ports/xnnpack/use-packages.patch
@ -1,116 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 2ca06ee..b0d8873 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -6526,7 +6526,7 @@ ELSE()
-   SET_PROPERTY(SOURCE ${COLD_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS "$<$<NOT:$<CONFIG:Debug>>: -Os >")
- ENDIF()
- 
-TARGET_INCLUDE_DIRECTORIES(XNNPACK PUBLIC include)
-+TARGET_INCLUDE_DIRECTORIES(XNNPACK PUBLIC $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include> $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
- TARGET_INCLUDE_DIRECTORIES(XNNPACK PRIVATE src)
- TARGET_INCLUDE_DIRECTORIES(all_microkernels PRIVATE include src)
- TARGET_INCLUDE_DIRECTORIES(packing PRIVATE include src)
-@@ -6547,7 +6547,7 @@ IF(LIBM)
- ENDIF()
- 
- # ---[ Configure clog
-IF(NOT TARGET clog)
-+IF(FALSE)
-   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
-     SET(CLOG_BUILD_TESTS OFF CACHE BOOL "")
-     SET(CLOG_RUNTIME_TYPE "${CPUINFO_RUNTIME_TYPE}" CACHE STRING "")
-@@ -6565,11 +6565,9 @@ IF(NOT TARGET clog)
-     SET_PROPERTY(TARGET clog PROPERTY IMPORTED_LOCATION "${CLOG_LIBRARY}")
-   ENDIF()
- ENDIF()
-TARGET_LINK_LIBRARIES(XNNPACK PRIVATE clog)
-TARGET_LINK_LIBRARIES(jit PRIVATE clog)
- 
- # ---[ Configure cpuinfo
-IF(NOT TARGET cpuinfo)
-+IF(false)
-   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
-     SET(CPUINFO_BUILD_TOOLS OFF CACHE BOOL "")
-     SET(CPUINFO_BUILD_UNIT_TESTS OFF CACHE BOOL "")
-@@ -6587,10 +6585,12 @@ IF(NOT TARGET cpuinfo)
-     SET_PROPERTY(TARGET cpuinfo PROPERTY IMPORTED_LOCATION "${CPUINFO_LIBRARY}")
-   ENDIF()
- ENDIF()
-TARGET_LINK_LIBRARIES(XNNPACK PRIVATE cpuinfo)
-+find_package(cpuinfo CONFIG REQUIRED)
-+TARGET_LINK_LIBRARIES(XNNPACK PRIVATE cpuinfo::clog cpuinfo::cpuinfo)
-+TARGET_LINK_LIBRARIES(jit PRIVATE cpuinfo::clog)
- 
- # ---[ Configure pthreadpool
-IF(NOT TARGET pthreadpool)
-+IF(FALSE)
-   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
-     SET(PTHREADPOOL_BUILD_TESTS OFF CACHE BOOL "")
-     SET(PTHREADPOOL_BUILD_BENCHMARKS OFF CACHE BOOL "")
-@@ -6607,14 +6607,15 @@ IF(NOT TARGET pthreadpool)
-     SET_PROPERTY(TARGET pthreadpool PROPERTY IMPORTED_LOCATION "${PTHREADPOOL_LIBRARY}")
-   ENDIF()
- ENDIF()
-TARGET_LINK_LIBRARIES(XNNPACK PUBLIC pthreadpool)
-TARGET_LINK_LIBRARIES(all_microkernels PRIVATE pthreadpool)
-TARGET_LINK_LIBRARIES(indirection PRIVATE pthreadpool)
-TARGET_LINK_LIBRARIES(jit PRIVATE pthreadpool)
-TARGET_LINK_LIBRARIES(packing PRIVATE pthreadpool)
-+find_package(unofficial-pthreadpool CONFIG REQUIRED)
-+target_link_libraries(XNNPACK PUBLIC unofficial::pthreadpool)
-+target_link_libraries(all_microkernels PRIVATE unofficial::pthreadpool)
-+target_link_libraries(indirection PRIVATE unofficial::pthreadpool)
-+target_link_libraries(jit PRIVATE unofficial::pthreadpool)
-+target_link_libraries(packing PRIVATE unofficial::pthreadpool)
- 
- # ---[ Configure FXdiv
-IF(NOT TARGET fxdiv)
-+IF(FALSE)
-   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
-     SET(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
-     SET(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
-@@ -6630,12 +6631,13 @@ IF(NOT TARGET fxdiv)
-     SET_PROPERTY(TARGET fxdiv PROPERTY LINKER_LANGUAGE C)
-   ENDIF()
- ENDIF()
-TARGET_LINK_LIBRARIES(XNNPACK PRIVATE fxdiv)
-TARGET_LINK_LIBRARIES(all_microkernels PRIVATE fxdiv)
-TARGET_LINK_LIBRARIES(indirection PRIVATE fxdiv)
-+find_path(FXDIV_INCLUDE_DIRS "fxdiv.h" REQUIRED)
-+target_include_directories(XNNPACK PRIVATE ${FXDIV_INCLUDE_DIRS})
-+target_include_directories(all_microkernels PRIVATE ${FXDIV_INCLUDE_DIRS})
-+target_include_directories(indirection PRIVATE ${FXDIV_INCLUDE_DIRS})
- 
- # ---[ Configure FP16
-IF(NOT TARGET fp16)
-+IF(FALSE)
-   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
-     SET(FP16_BUILD_TESTS OFF CACHE BOOL "")
-     SET(FP16_BUILD_BENCHMARKS OFF CACHE BOOL "")
-@@ -6651,15 +6653,21 @@ IF(NOT TARGET fp16)
-     SET_PROPERTY(TARGET fp16 PROPERTY LINKER_LANGUAGE C)
-   ENDIF()
- ENDIF()
-TARGET_LINK_LIBRARIES(XNNPACK PRIVATE fp16)
-TARGET_LINK_LIBRARIES(all_microkernels PRIVATE fp16)
-TARGET_LINK_LIBRARIES(packing PRIVATE fp16)
-TARGET_LINK_LIBRARIES(indirection PRIVATE fp16)
-+find_path(FP16_INCLUDE_DIRS "fp16.h" REQUIRED)
-+target_include_directories(XNNPACK PRIVATE ${FP16_INCLUDE_DIRS})
-+target_include_directories(all_microkernels PRIVATE ${FP16_INCLUDE_DIRS})
-+target_include_directories(packing PRIVATE ${FP16_INCLUDE_DIRS})
-+target_include_directories(indirection PRIVATE ${FP16_INCLUDE_DIRS})
- 
- INSTALL(TARGETS XNNPACK
-+    EXPORT xnnpack-config
-+    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
-     LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
-     ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
-     PUBLIC_HEADER DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
-+install(EXPORT xnnpack-config
-+    NAMESPACE unofficial::
-+    DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/xnnpack)
- 
- # ---[ XNNPACK unit tests
- IF(XNNPACK_BUILD_TESTS)
--- a/ports/xnnpack/vcpkg.json
+++ b/ports/xnnpack/vcpkg.json
@ -1,10 +1,10 @@
 {
  "name": "xnnpack",
-  "version-date": "2022-02-17",
+  "version-date": "2022-12-22",
  "description": "High-efficiency floating-point neural network inference operators for mobile, server, and Web",
  "homepage": "https://github.com/google/XNNPACK",
  "license": "BSD-3-Clause",
-  "supports": "!(arm & windows) & !uwp",
+  "supports": "!(arm & windows) & !uwp & !arm32",
  "dependencies": [
    "cpuinfo",
    "fp16",
@ -13,10 +13,6 @@
    {
      "name": "vcpkg-cmake",
      "host": true
-    },
-    {
-      "name": "vcpkg-cmake-config",
-      "host": true
    }
  ]
 }
--- a/scripts/ci.baseline.txt
+++ b/scripts/ci.baseline.txt
@ -627,10 +627,6 @@ libtcod:arm64-android=fail
 libtcod:x64-android=fail
 libtomcrypt:arm64-windows=fail
 libtomcrypt:arm64-uwp=fail
-# D:\b\libtorch\src\v1.12.1-e1cb316ec6.clean\aten\src\ATen/cpu/vec/vec_base.h(134): fatal error C1001: Internal compiler error.
-# (compiler file 'D:\a\_work\1\s\src\vctools\Compiler\CxxFE\sl\p1\c\toinil.c', line 910)
-libtorch:x64-windows=fail
-libtorch:x64-osx=fail
 libusb-win32:arm64-uwp=fail
 libusb-win32:x64-uwp=fail
 #Skip detection to avoid upstream remove older releases
--- a/versions/baseline.json
+++ b/versions/baseline.json
@ -4949,8 +4949,8 @@
      "port-version": 0
    },
    "libtorch": {
-      "baseline": "1.12.1",
-      "port-version": 5
+      "baseline": "2.1.2",
+      "port-version": 0
    },
    "libtorrent": {
      "baseline": "2.0.9",
@ -9333,7 +9333,7 @@
      "port-version": 2
    },
    "xnnpack": {
-      "baseline": "2022-02-17",
+      "baseline": "2022-12-22",
      "port-version": 0
    },
    "xorg-macros": {
--- a/versions/l-/libtorch.json
+++ b/versions/l-/libtorch.json
@ -1,5 +1,10 @@
 {
  "versions": [
+    {
+      "git-tree": "418db4a2514c9530b67097fca0ce7cafde5a2a15",
+      "version": "2.1.2",
+      "port-version": 0
+    },
    {
      "git-tree": "20a2f323fb33dab29315fd18217e263ade2a9bd6",
      "version": "1.12.1",
--- a/versions/x-/xnnpack.json
+++ b/versions/x-/xnnpack.json
@ -1,5 +1,10 @@
 {
  "versions": [
+    {
+      "git-tree": "7a84132ceac551e5ba3da8117fdc23f4f1571fd2",
+      "version-date": "2022-12-22",
+      "port-version": 0
+    },
    {
      "git-tree": "439631254e564280e520c848b6b16e259c79b422",
      "version-date": "2022-02-17",