[libtorch] [xnnpack] [qnnpack] [cpuinfo][fbgemm] (#46649)

Co-authored-by: Steve <steve@imagus.au> Co-authored-by: Steve Brain <sbrain@vixvizion.com> Co-authored-by: Kai Pastor <dg0yt@darc.de> Co-authored-by: luncliff <luncliff@gmail.com>
2025-09-20 03:27:54 +10:00 · 2025-09-20 03:27:54 +10:00 · 29ff5b8131
parent cc3290ef83
commit 29ff5b8131
43 changed files with 1120 additions and 1682 deletions
--- a/ports/cpuinfo/add-clog-cmake.patch
+++ b/ports/cpuinfo/add-clog-cmake.patch
@ -0,0 +1,13 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index b260162..d63e180 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -929,3 +929,7 @@ IF(CPUINFO_BUILD_PKG_CONFIG)
+   DESTINATION "${CMAKE_INSTALL_LIBDIR}/pkgconfig")
+ 
+ ENDIF()
+
+IF(CPUINFO_BUILD_CLOG)
+add_subdirectory(deps/clog)
+ENDIF()
+\ No newline at end of file
--- a/ports/cpuinfo/portfile.cmake
+++ b/ports/cpuinfo/portfile.cmake
@ -6,14 +6,17 @@ endif()
 vcpkg_from_github(
    OUT_SOURCE_PATH SOURCE_PATH
    REPO pytorch/cpuinfo
-    REF 5e63739504f0f8e18e941bd63b2d6d42536c7d90
-    SHA512 6a61f4574661a55771c2ec31bb0919a51d0bd8c770477b254a5c14dc5323716af275c7fe3abc5aa96720d7cc929559ca66f614265d3940e076b8db2fa15c8e36
+    REF 33ed0be77d7767d0e2010e2c3cf972ef36c7c307
+    SHA512 41775802b5a33f1bb0cb2c5dd126180e25d281a57c934d94247130c31ec7bf1dbf269ed9e8e946be1bdeeb09f0c17fee42048f71b4beb44c65df18cfb0cddc3b
    HEAD_REF master
+    PATCHES
+        add-clog-cmake.patch
 )

 vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS
    FEATURES
        tools CPUINFO_BUILD_TOOLS
+        clog CPUINFO_BUILD_CLOG
 )

 set(LINK_OPTIONS "")
@ -37,6 +40,7 @@ vcpkg_cmake_configure(
        -DCPUINFO_BUILD_UNIT_TESTS=OFF
        -DCPUINFO_BUILD_MOCK_TESTS=OFF
        -DCPUINFO_BUILD_BENCHMARKS=OFF
+        -DCLOG_BUILD_TESTS=OFF
    OPTIONS_DEBUG
        -DCPUINFO_LOG_LEVEL=debug
    OPTIONS_RELEASE
--- a/ports/cpuinfo/vcpkg.json
+++ b/ports/cpuinfo/vcpkg.json
@ -1,7 +1,6 @@
 {
  "name": "cpuinfo",
-  "version-date": "2022-07-19",
-  "port-version": 3,
+  "version-date": "2025-07-25",
  "description": "CPU INFOrmation library (x86/x86-64/ARM/ARM64, Linux/Windows/Android/macOS/iOS)",
  "homepage": "https://github.com/pytorch/cpuinfo",
  "license": "BSD-2-Clause",
@ -17,6 +16,10 @@
    }
  ],
  "features": {
+    "clog": {
+      "description": "Build clog support",
+      "supports": "!uwp"
+    },
    "tools": {
      "description": "Build cpuinfo command-line tools",
      "supports": "!uwp"
--- a/ports/fbgemm/fix-cmakelists.patch
+++ b/ports/fbgemm/fix-cmakelists.patch
@ -1,5 +1,5 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 0243f68..ccf7331 100644
+index 0243f68..bbe98e9 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
@@ -94,13 +94,12 @@ project(fbgemm VERSION 0.1 LANGUAGES CXX C)
@ -21,7 +21,7 @@ index 0243f68..ccf7331 100644
 option(FBGEMM_BUILD_TESTS "Build fbgemm unit tests" ON)
 option(FBGEMM_BUILD_BENCHMARKS "Build fbgemm benchmarks" ON)
 option(FBGEMM_BUILD_DOCS "Build fbgemm documentation" OFF)
-@@ -178,7 +177,7 @@ add_dependencies(fbgemm_autovec defs.bzl)
+@@ -178,11 +177,12 @@ add_dependencies(fbgemm_autovec defs.bzl)
 # 2)  MSVC uses /MD in default cxx compiling flags,
 # Need to change it to /MT in static case
 if(MSVC)
@ -30,7 +30,20 @@ index 0243f68..ccf7331 100644
   if(FBGEMM_LIBRARY_TYPE STREQUAL "static")
     target_compile_definitions(fbgemm_generic PRIVATE ASMJIT_STATIC)
     target_compile_definitions(fbgemm_avx2 PRIVATE ASMJIT_STATIC)
-@@ -199,7 +198,6 @@ if(MSVC)
+     target_compile_definitions(fbgemm_avx512 PRIVATE ASMJIT_STATIC)
+if(FALSE)
+     foreach(flag_var
+       CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+       CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+@@ -190,6 +190,7 @@ if(MSVC)
+         string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
+       endif(${flag_var} MATCHES "/MD")
+     endforeach(flag_var)
+endif()    
+   endif()
+   target_compile_options(fbgemm_avx2 PRIVATE "/arch:AVX2")
+   target_compile_options(fbgemm_avx512 PRIVATE "/arch:AVX512")
+@@ -199,7 +200,6 @@ if(MSVC)
 else(MSVC)
   string(APPEND CMAKE_CXX_FLAGS " -Wall")
   string(APPEND CMAKE_CXX_FLAGS " -Wextra")
@ -38,27 +51,25 @@ index 0243f68..ccf7331 100644
   string(APPEND CMAKE_CXX_FLAGS " -Wno-deprecated-declarations")
   string(APPEND CMAKE_CXX_FLAGS " -Wunknown-pragmas")
   string(APPEND CMAKE_CXX_FLAGS " -Wimplicit-fallthrough")
-@@ -247,7 +245,8 @@ message(WARNING "CMAKE_CXX_FLAGS_DEBUG is ${CMAKE_CXX_FLAGS_DEBUG}")
+@@ -247,6 +247,8 @@ message(WARNING "CMAKE_CXX_FLAGS_DEBUG is ${CMAKE_CXX_FLAGS_DEBUG}")
 message(WARNING "CMAKE_CXX_FLAGS_RELEASE is ${CMAKE_CXX_FLAGS_RELEASE}")
 message(WARNING "==========")
 
-if(NOT TARGET asmjit)
 +find_package(asmjit CONFIG REQUIRED) # target 'asmjit::asmjit'
-+if(FALSE)
+add_library(asmjit ALIAS asmjit::asmjit)
+ if(NOT TARGET asmjit)
   # Download asmjit from github if ASMJIT_SRC_DIR is not specified.
   if(NOT DEFINED ASMJIT_SRC_DIR)
-     set(ASMJIT_SRC_DIR "${FBGEMM_SOURCE_DIR}/external/asmjit"
-@@ -280,7 +279,8 @@ if(NOT TARGET asmjit)
+@@ -280,6 +282,8 @@ if(NOT TARGET asmjit)
   endif()
 endif()
 
-if(NOT TARGET cpuinfo)
 +find_package(cpuinfo CONFIG REQUIRED) # target 'cpuinfo::cpuinfo'
-+if(FALSE)
+add_library(cpuinfo ALIAS cpuinfo::cpuinfo)
+ if(NOT TARGET cpuinfo)
   #Download cpuinfo from github if CPUINFO_SOURCE_DIR is not specified.
   if(NOT DEFINED CPUINFO_SOURCE_DIR)
-     set(CPUINFO_SOURCE_DIR "${FBGEMM_SOURCE_DIR}/external/cpuinfo"
-@@ -301,26 +301,26 @@ endif()
+@@ -301,26 +305,26 @@ endif()
 target_include_directories(fbgemm_generic BEFORE
       PUBLIC $<BUILD_INTERFACE:${FBGEMM_SOURCE_DIR}>
       PUBLIC $<BUILD_INTERFACE:${FBGEMM_SOURCE_DIR}/include>
@ -93,7 +104,7 @@ index 0243f68..ccf7331 100644
 
 if(FBGEMM_LIBRARY_TYPE STREQUAL "default")
   add_library(fbgemm
-@@ -338,6 +338,7 @@ elseif(FBGEMM_LIBRARY_TYPE STREQUAL "shared")
+@@ -338,6 +342,7 @@ elseif(FBGEMM_LIBRARY_TYPE STREQUAL "shared")
   set_property(TARGET fbgemm_avx2 PROPERTY POSITION_INDEPENDENT_CODE ON)
   set_property(TARGET fbgemm_avx512 PROPERTY POSITION_INDEPENDENT_CODE ON)
   set_property(TARGET fbgemm_autovec PROPERTY POSITION_INDEPENDENT_CODE ON)
@ -101,7 +112,24 @@ index 0243f68..ccf7331 100644
 elseif(FBGEMM_LIBRARY_TYPE STREQUAL "static")
   add_library(fbgemm STATIC
     $<TARGET_OBJECTS:fbgemm_generic>
-@@ -364,15 +365,10 @@ target_include_directories(fbgemm BEFORE
+@@ -346,11 +351,11 @@ elseif(FBGEMM_LIBRARY_TYPE STREQUAL "static")
+     $<TARGET_OBJECTS:fbgemm_autovec>)
+   #MSVC need to define FBGEMM_STATIC for fbgemm_generic also to
+   #avoid generating _dllimport functions.
+-  target_compile_definitions(fbgemm_generic PRIVATE FBGEMM_STATIC)
+-  target_compile_definitions(fbgemm_avx2 PRIVATE FBGEMM_STATIC)
+-  target_compile_definitions(fbgemm_avx512 PRIVATE FBGEMM_STATIC)
+-  target_compile_definitions(fbgemm_autovec PRIVATE FBGEMM_STATIC)
+-  target_compile_definitions(fbgemm PRIVATE FBGEMM_STATIC)
+  target_compile_definitions(fbgemm_generic PUBLIC FBGEMM_STATIC)
+  target_compile_definitions(fbgemm_avx2 PUBLIC FBGEMM_STATIC)
+  target_compile_definitions(fbgemm_avx512 PUBLIC FBGEMM_STATIC)
+  target_compile_definitions(fbgemm_autovec PUBLIC FBGEMM_STATIC)
+  target_compile_definitions(fbgemm PUBLIC FBGEMM_STATIC)
+ else()
+   message(FATAL_ERROR "Unsupported library type ${FBGEMM_LIBRARY_TYPE}")
+ endif()
+@@ -364,20 +369,15 @@ target_include_directories(fbgemm BEFORE
     PUBLIC $<BUILD_INTERFACE:${FBGEMM_SOURCE_DIR}>
     PUBLIC $<BUILD_INTERFACE:${FBGEMM_SOURCE_DIR}/include>)
 
@ -119,12 +147,44 @@ index 0243f68..ccf7331 100644
 endif()
 
 install(
-@@ -391,7 +387,7 @@ install(
-   DESTINATION share/cmake/fbgemm
-   FILE fbgemmLibraryConfig.cmake)
+   TARGETS fbgemm
+-  EXPORT fbgemmLibraryConfig
+  EXPORT unofficial-fbgemm-config-targets
+   ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+   LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+   RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}) #For windows
+@@ -387,11 +387,16 @@ install(
+   DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/fbgemm")
+ 
+ install(
+-  EXPORT fbgemmLibraryConfig
+-  DESTINATION share/cmake/fbgemm
+-  FILE fbgemmLibraryConfig.cmake)
+  EXPORT unofficial-fbgemm-config-targets
+  FILE unofficial-fbgemm-config-targets.cmake
+  DESTINATION  ${CMAKE_INSTALL_DATAROOTDIR}/fbgemmLibrary)
 
 -if(MSVC)
+include(CMakePackageConfigHelpers)
+configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in "${CMAKE_CURRENT_BINARY_DIR}/fbgemmLibraryConfig.cmake" INSTALL_DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/fbgemmLibrary)
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/fbgemmLibraryConfig.cmake DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/fbgemmLibrary)
+
+
 +if(FALSE)
   if(FBGEMM_LIBRARY_TYPE STREQUAL "shared")
     install(
       FILES $<TARGET_PDB_FILE:fbgemm> $<TARGET_PDB_FILE:asmjit>
+diff --git a/Config.cmake.in b/Config.cmake.in
+new file mode 100644
+index 0000000..e2e187d
+--- /dev/null
+++ b/Config.cmake.in
+@@ -0,0 +1,8 @@
+@PACKAGE_INIT@
+
+include(CMakeFindDependencyMacro)
+find_dependency(asmjit)
+find_dependency(cpuinfo)
+
+
+include ( "${CMAKE_CURRENT_LIST_DIR}/unofficial-fbgemm-config-targets.cmake" )
--- a/ports/fbgemm/portfile.cmake
+++ b/ports/fbgemm/portfile.cmake
@ -20,7 +20,7 @@ vcpkg_cmake_configure(
 )
 vcpkg_cmake_install()
 vcpkg_copy_pdbs()
-vcpkg_cmake_config_fixup(PACKAGE_NAME fbgemmLibrary CONFIG_PATH share/cmake/${PORT})
+vcpkg_cmake_config_fixup(PACKAGE_NAME fbgemmLibrary)

 # this internal header is required by pytorch
 file(INSTALL     "${SOURCE_PATH}/src/RefImplementations.h"
--- a/ports/fbgemm/vcpkg.json
+++ b/ports/fbgemm/vcpkg.json
@ -1,6 +1,7 @@
 {
  "name": "fbgemm",
  "version": "1.0.0",
+  "port-version": 1,
  "description": "FB (Facebook) + GEMM (General Matrix-Matrix Multiplication)",
  "homepage": "https://code.fb.com/ml-applications/fbgemm/",
  "license": "BSD-3-Clause",
--- a/ports/libtorch/add-include-chrono.patch
+++ b/ports/libtorch/add-include-chrono.patch
@ -1,11 +0,0 @@
-diff --git a/torch/csrc/jit/runtime/logging.h b/torch/csrc/jit/runtime/logging.h
-index 5499ecf..269168e 100644
--- a/torch/csrc/jit/runtime/logging.h
-+++ b/torch/csrc/jit/runtime/logging.h
-@@ -1,5 +1,6 @@
- #pragma once
- 
-+#include <chrono>
- #include <mutex>
- #include <string>
- #include <unordered_map>
--- a/ports/libtorch/clang-cl.patch
+++ b/ports/libtorch/clang-cl.patch
@ -1,47 +0,0 @@
-diff --git a/c10/util/Exception.h b/c10/util/Exception.h
-index 2dd1001..b3f3d60 100644
--- a/c10/util/Exception.h
-+++ b/c10/util/Exception.h
-@@ -233,12 +233,16 @@ struct C10_API WarnAlways {
- // IndexError when they cross to Python.
- class C10_API IndexError : public Error {
-   using Error::Error;
-+public:
-+  inline IndexError(SourceLocation source_location, std::string msg) : Error(source_location,msg) {};
- };
- 
- // Used in ATen for invalid values.  These turn into
- // ValueError when they cross to Python.
- class C10_API ValueError : public Error {
-   using Error::Error;
-+public:
-+  inline ValueError(SourceLocation source_location, std::string msg) : Error(source_location,msg) {};
- };
- 
- // Used in ATen for invalid types.  These turn into
-diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
-index cf349c2..f7a595e 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -830,7 +830,7 @@ if(BUILD_CAFFE2 AND NOT MSVC)
-   target_compile_options(torch_cpu PRIVATE "-Wno-sign-compare")
- endif()
- 
-if("${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
@ -1,254 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 3a48eaf..7b8bc7c 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 fd6b7ff..d185233 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 feebad7..ad407e1 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 c3abce5..63f665f 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)
- 
-@@ -1127,7 +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()
-@@ -1231,7 +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}")
-@@ -1244,8 +1247,10 @@ 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()
- 
- # ---[ HIP
-@@ -1555,7 +1560,10 @@ 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)
-@@ -1573,23 +1581,13 @@ 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)
-    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)
-    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}")
-    list(APPEND Caffe2_DEPENDENCY_LIBS onnx_proto onnx)
-+    find_package(ONNX CONFIG REQUIRED)
-+    set(ONNX_LIBRARY ONNX::onnx)
-+    set(ONNX_PROTO_LIBRARY ONNX::onnx_proto)
-+    list(APPEND Caffe2_DEPENDENCY_LIBS ${ONNX_LIBRARY} ${ONNX_PROTO_LIBRARY})
-   endif()
-   include_directories(${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 +1832,8 @@ 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(${PROJECT_SOURCE_DIR}/third_party/fmt)
-+find_package(fmt REQUIRED)
- # Disable compiler feature checks for `fmt`.
- #
- # CMake compiles a little program to check compiler features. Some of our build
-@@ -1843,7 +1841,6 @@ add_subdirectory(${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 ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
-diff --git a/cmake/Modules/FindLMDB.cmake b/cmake/Modules/FindLMDB.cmake
-index 2f0adb1..8a817fd 100644
--- a/cmake/Modules/FindLMDB.cmake
-+++ b/cmake/Modules/FindLMDB.cmake
-@@ -12,12 +12,8 @@
- # Copyright 2013 Conrad Steenberg <conrad.steenberg@gmail.com>
- # Aug 31, 2013
- 
-if(MSVC)
-  find_package(LMDB NO_MODULE)
-else()
-  find_path(LMDB_INCLUDE_DIR NAMES  lmdb.h PATHS "$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 "$ENV{LMDB_DIR}/lib" )
- 
- include(FindPackageHandleStandardArgs)
- find_package_handle_standard_args(LMDB DEFAULT_MSG LMDB_INCLUDE_DIR LMDB_LIBRARIES)
-diff --git a/cmake/public/cuda.cmake b/cmake/public/cuda.cmake
-index 32f3ba3..4112937 100644
--- a/cmake/public/cuda.cmake
-+++ b/cmake/public/cuda.cmake
-@@ -67,8 +67,8 @@ if(NOT CMAKE_CUDA_COMPILER_VERSION STREQUAL CUDAToolkit_VERSION OR
-                       "V${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 4d48c0f..ebdd39a 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}" "${value}")
-   endif()
- endmacro()
- 
--- a/ports/libtorch/cuda-adjustments.patch
+++ b/ports/libtorch/cuda-adjustments.patch
@ -1,63 +0,0 @@
-diff --git a/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake b/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
-index 420ee63..8bc24a0 100644
--- a/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
-+++ b/cmake/Modules_CUDA_fix/upstream/FindCUDA.cmake
-@@ -1691,14 +1691,6 @@ function(CUDA_LINK_SEPARABLE_COMPILATION_OBJECTS output_file cuda_target options
-     # If -ccbin, --compiler-bindir has been specified, don't do anything.  Otherwise add it here.
-     list( FIND nvcc_flags "-ccbin" ccbin_found0 )
-     list( FIND nvcc_flags "--compiler-bindir" ccbin_found1 )
-    if( ccbin_found0 LESS 0 AND ccbin_found1 LESS 0 AND CUDA_HOST_COMPILER )
-      # Match VERBATIM check below.
-      if(CUDA_HOST_COMPILER MATCHES "\\$\\(VCInstallDir\\)")
-        list(APPEND nvcc_flags -ccbin "\"${CUDA_HOST_COMPILER}\"")
-      else()
-        list(APPEND nvcc_flags -ccbin "${CUDA_HOST_COMPILER}")
-      endif()
-    endif()
- 
-     # Create a list of flags specified by CUDA_NVCC_FLAGS_${CONFIG} and CMAKE_${CUDA_C_OR_CXX}_FLAGS*
-     set(config_specific_flags)
-diff --git a/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake b/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
-index 9293df3..8c12015 100644
--- a/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
-+++ b/cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
-@@ -126,13 +126,6 @@ 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 (CUDA_HOST_COMPILER STREQUAL "@_CUDA_MSVC_HOST_COMPILER@" AND DEFINED CCBIN)
-    set(CCBIN -ccbin "${CCBIN}")
-  else()
-    set(CCBIN -ccbin "${CUDA_HOST_COMPILER}")
-  endif()
-endif()
- 
- # cuda_execute_process - Executes a command with optional command echo and status message.
- #
-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..6d0d4fa 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 +30,6 @@ 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")
- 
- # This list is used to filter CUDA archs when autodetecting
- set(CUDA_ALL_GPU_ARCHITECTURES "3.5" "5.0")
-diff --git a/cmake/public/cuda.cmake b/cmake/public/cuda.cmake
-index 4112937..5964429 100644
--- a/cmake/public/cuda.cmake
-+++ b/cmake/public/cuda.cmake
-@@ -41,9 +41,6 @@ set(CUDAToolkit_ROOT "${CUDA_TOOLKIT_ROOT_DIR}")
- # Pass clang as host compiler, which according to the docs
- # Must be done before CUDA language is enabled, see
- # https://cmake.org/cmake/help/v3.15/variable/CMAKE_CUDA_HOST_COMPILER.html
-if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
-  set(CMAKE_CUDA_HOST_COMPILER "${CMAKE_C_COMPILER}")
-endif()
- enable_language(CUDA)
- if("X${CMAKE_CUDA_STANDARD}" STREQUAL "X" )
-   set(CMAKE_CUDA_STANDARD ${CMAKE_CXX_STANDARD})
--- a/ports/libtorch/fix-api-export.patch
+++ b/ports/libtorch/fix-api-export.patch
@ -1,39 +0,0 @@
-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-aten-cutlass.patch
+++ b/ports/libtorch/fix-aten-cutlass.patch
@ -1,32 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 80bbfc6b..eead313d 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -560,6 +560,8 @@ if(MSVC)
- 
-   string(APPEND CMAKE_CXX_FLAGS " /FS")
-   string(APPEND CMAKE_CUDA_FLAGS " -Xcompiler /FS")
-+  # CUTLASS_CONSTEXPR_IF_CXX17 must be constexpr. Correct the __cplusplus value
-+  string(APPEND CMAKE_CUDA_FLAGS " -Xcompiler /Zc:__cplusplus")
- endif(MSVC)
- 
- string(APPEND CMAKE_CUDA_FLAGS " -Xfatbin -compress-all")
-diff --git a/aten/src/ATen/CMakeLists.txt b/aten/src/ATen/CMakeLists.txt
-index 93021a74..0755c890 100644
--- a/aten/src/ATen/CMakeLists.txt
-+++ b/aten/src/ATen/CMakeLists.txt
-@@ -437,7 +437,13 @@ if(NOT MSVC AND NOT EMSCRIPTEN AND NOT INTERN_BUILD_MOBILE)
- endif()
- 
- if(USE_CUDA AND NOT USE_ROCM)
-  list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
-+  find_package(NvidiaCutlass CONFIG REQUIRED)
-+  get_target_property(CUTLASS_INCLUDE_DIRS nvidia::cutlass::cutlass INTERFACE_INCLUDE_DIRECTORIES)
-+  list(APPEND ATen_CUDA_INCLUDE ${CUTLASS_INCLUDE_DIRS})
-+  if(MSVC)
-+    # CUTLASS_CONSTEXPR_IF_CXX17 must be constexpr. Correct the __cplusplus value with MSVC
-+    add_compile_options($<$<COMPILE_LANGUAGE:CUDA>:"-Xcompiler /Zc:__cplusplus">)
-+  endif()
-   if($ENV{ATEN_STATIC_CUDA})
-     list(APPEND ATen_CUDA_DEPENDENCY_LIBS
-       ${CUDA_LIBRARIES}
--- a/ports/libtorch/fix-build-error-with-fmt11.patch
+++ b/ports/libtorch/fix-build-error-with-fmt11.patch
@ -1,36 +0,0 @@
-diff --git a/torch/csrc/distributed/c10d/socket.cpp b/torch/csrc/distributed/c10d/socket.cpp
-index 51ea05fa..3c3fbb39 100644
--- a/torch/csrc/distributed/c10d/socket.cpp
-+++ b/torch/csrc/distributed/c10d/socket.cpp
-@@ -29,6 +29,7 @@
- 
- #include <fmt/chrono.h>
- #include <fmt/format.h>
-+#include <fmt/ranges.h>
- 
- #include <torch/csrc/distributed/c10d/error.h>
- #include <torch/csrc/distributed/c10d/exception.h>
-diff --git a/torch/csrc/profiler/standalone/execution_trace_observer.cpp b/torch/csrc/profiler/standalone/execution_trace_observer.cpp
-index 93d9e2d1..ea241a8b 100644
--- a/torch/csrc/profiler/standalone/execution_trace_observer.cpp
-+++ b/torch/csrc/profiler/standalone/execution_trace_observer.cpp
-@@ -10,6 +10,7 @@
- #endif // _WIN32
- 
- #include <fmt/format.h>
-+#include <fmt/ranges.h>
- #include <algorithm>
- #include <chrono>
- #include <cmath>
-diff --git a/torch/csrc/profiler/util.cpp b/torch/csrc/profiler/util.cpp
-index 180555f9..1358c9ab 100644
--- a/torch/csrc/profiler/util.cpp
-+++ b/torch/csrc/profiler/util.cpp
-@@ -5,6 +5,7 @@
- #include <c10/util/ArrayRef.h>
- #include <c10/util/irange.h>
- #include <fmt/format.h>
-+#include <fmt/ranges.h>
- 
- #ifdef USE_KINETO
- #include <libkineto.h>
--- a/ports/libtorch/fix-build.patch
+++ b/ports/libtorch/fix-build.patch
@ -1,156 +0,0 @@
-diff --git a/c10/cuda/CMakeLists.txt b/c10/cuda/CMakeLists.txt
-index c0628d0..5dfb7f2 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 ecbc7c1..2dd1001 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/pyproject.toml b/pyproject.toml
-index eb764cb..c70f317 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 7016441..a6416c7 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 cd57825..75e81e9 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 d87e97a..6c52f39 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 20158a9..7d3c475 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 ${LIBSHM_INSTALL_LIB_SUBDIR})
-+install(TARGETS shm 
-+  RUNTIME DESTINATION bin
-+  ARCHIVE DESTINATION lib
-+  LIBRARY DESTINATION lib
-+)
- install(FILES libshm.h DESTINATION "include")
- install(TARGETS torch_shm_manager DESTINATION "bin")
-diff --git a/torch/lib/libshm_windows/CMakeLists.txt b/torch/lib/libshm_windows/CMakeLists.txt
-index df2a106..b86e0b6 100644
--- a/torch/lib/libshm_windows/CMakeLists.txt
-+++ b/torch/lib/libshm_windows/CMakeLists.txt
-@@ -19,9 +19,12 @@ target_include_directories(shm PRIVATE
- target_link_libraries(shm torch c10)
- 
- 
-install(TARGETS shm DESTINATION "${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 6600eb3..15ddc8c 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-calculate-minloglevel.patch
+++ b/ports/libtorch/fix-calculate-minloglevel.patch
@ -1,13 +0,0 @@
-diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
-index 306dac5..dca802c 100644
--- a/c10/util/Logging.cpp
-+++ b/c10/util/Logging.cpp
-@@ -272,7 +272,7 @@ void UpdateLoggingLevelsFromFlags() {
- 
- void ShowLogInfoToStderr() {
-   FLAGS_logtostderr = 1;
-  FLAGS_minloglevel = std::min(FLAGS_minloglevel, google::GLOG_INFO);
-+  FLAGS_minloglevel = std::min(FLAGS_minloglevel, static_cast<decltype(FLAGS_minloglevel)>(google::GLOG_INFO));
- }
- } // namespace c10
- 
--- a/ports/libtorch/fix-cmake.patch
+++ b/ports/libtorch/fix-cmake.patch
@ -0,0 +1,404 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 07edb30..1ca26e3 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -1247,8 +1247,7 @@ 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 CONFIG REQUIRED)
+ endif()
+ 
+ if(USE_MIMALLOC AND USE_MIMALLOC_ON_MKL)
+@@ -1316,10 +1315,7 @@ if(BUILD_SHARED_LIBS)
+     DIRECTORY ${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/CMakeLists.txt b/aten/src/ATen/CMakeLists.txt
+index 085af37..1c61f27 100644
+--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
+@@ -515,7 +515,7 @@ if(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")
+@@ -523,7 +523,7 @@ if(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})
+ 
+   if(NOT MSVC)
+     set(CMAKE_C_FLAGS_DEBUG ${OLD_CMAKE_C_FLAGS_DEBUG})
+@@ -534,8 +534,8 @@ if(USE_CUDA AND NOT USE_ROCM)
+   add_definitions(-DCUTLASS_ENABLE_TENSOR_CORE_MMA=1)
+   add_definitions(-DCUTLASS_ENABLE_SM90_EXTENDED_MMA_SHAPES=1)
+   add_definitions(-DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED)
+-  list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
+-  list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
+  find_package(NvidiaCutlass CONFIG REQUIRED)
+  list(APPEND ATen_CUDA_DEPENDENCY_LIBS nvidia::cutlass::cutlass)
+   if($ENV{ATEN_STATIC_CUDA})
+     list(APPEND ATen_CUDA_DEPENDENCY_LIBS
+       ${CUDA_LIBRARIES}
+diff --git a/c10/CMakeLists.txt b/c10/CMakeLists.txt
+index 34577ca..5462cfe 100644
+--- a/c10/CMakeLists.txt
+++ b/c10/CMakeLists.txt
+@@ -121,8 +121,7 @@ if(NOT BUILD_LIBTORCHLESS)
+   endif()
+ 
+   if(USE_MIMALLOC)
+-    target_link_libraries(c10 PRIVATE "mimalloc-static")
+-    add_dependencies(c10 mimalloc-static)
+    target_link_libraries(c10 PRIVATE "mimalloc")
+   endif()
+ 
+   if(LINUX)
+@@ -163,7 +162,7 @@ if(NOT BUILD_LIBTORCHLESS)
+   # 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)
+ endif()
+ 
+ install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}
+@@ -173,5 +172,5 @@ install(FILES ${CMAKE_BINARY_DIR}/c10/macros/cmake_macros.h
+         DESTINATION include/c10/macros)
+ 
+ if(MSVC AND C10_BUILD_SHARED_LIBS)
+-  install(FILES $<TARGET_PDB_FILE:c10> DESTINATION lib OPTIONAL)
+  install(FILES $<TARGET_PDB_FILE:c10> DESTINATION bin OPTIONAL)
+ endif()
+diff --git a/c10/cuda/CMakeLists.txt b/c10/cuda/CMakeLists.txt
+index 3327dab..4a0476f 100644
+--- a/c10/cuda/CMakeLists.txt
+++ b/c10/cuda/CMakeLists.txt
+@@ -82,7 +82,7 @@ if(NOT BUILD_LIBTORCHLESS)
+ # Note: for now, we will put all export path into one single Caffe2Targets group
+ # to deal with the cmake deployment need. Inside the Caffe2Targets set, the
+ # individual libraries like libc10.so and libcaffe2.so are still self-contained.
+-install(TARGETS c10_cuda EXPORT Caffe2Targets DESTINATION lib)
+install(TARGETS c10_cuda EXPORT Caffe2Targets  RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}" LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}" ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ 
+ endif()
+ 
+diff --git a/caffe2/CMakeLists.txt b/caffe2/CMakeLists.txt
+index d2d23b7..51930dd 100644
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -86,7 +86,7 @@ endif()
+ # ---[ Caffe2 build
+ # Note: the folders that are being commented out have not been properly
+ # addressed yet.
+-
+if(FALSE)
+ if(NOT MSVC AND USE_XNNPACK)
+   if(NOT TARGET fxdiv)
+     set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+@@ -96,6 +96,7 @@ if(NOT MSVC AND USE_XNNPACK)
+       "${CMAKE_BINARY_DIR}/FXdiv")
+   endif()
+ endif()
+endif()
+ 
+ add_subdirectory(core)
+ add_subdirectory(serialize)
+@@ -557,7 +558,7 @@ if(USE_CUDA)
+   endif()
+ 
+   target_link_libraries(caffe2_nvrtc PRIVATE caffe2::nvrtc ${DELAY_LOAD_FLAGS})
+-  install(TARGETS caffe2_nvrtc DESTINATION "${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)
+@@ -1134,7 +1135,7 @@ if(USE_XPU)
+   endif()
+ endif()
+ 
+-if(NOT MSVC AND USE_XNNPACK)
+if(FALSE)
+   TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+ endif()
+ 
+@@ -1251,11 +1252,11 @@ if(USE_KINETO)
+     ${TORCH_ROOT}/third_party/kineto/libkineto/src)
+ endif()
+ 
+-target_include_directories(torch_cpu PRIVATE
+-  ${TORCH_ROOT}/third_party/cpp-httplib)
+target_link_libraries(torch_cpu PRIVATE httplib::httplib nlohmann_json::nlohmann_json)
+if(APPLE)
+  target_link_libraries(torch_cpu PUBLIC "-framework CoreFoundation" "-framework CFNetwork")
+ 
+-target_include_directories(torch_cpu PRIVATE
+-  ${TORCH_ROOT}/third_party/nlohmann/include)
+endif()
+ 
+ install(DIRECTORY "${TORCH_SRC_DIR}/csrc"
+   DESTINATION ${TORCH_INSTALL_INCLUDE_DIR}/torch
+@@ -1536,17 +1537,17 @@ endif()
+ 
+ caffe2_interface_library(torch torch_library)
+ 
+-install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets DESTINATION "${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 "${TORCH_INSTALL_LIB_DIR}")
+ elseif(USE_XPU)
+-  install(TARGETS torch_xpu torch_xpu_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS torch_xpu torch_xpu_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 "${TORCH_INSTALL_LIB_DIR}")
+ 
+ target_link_libraries(torch PUBLIC torch_cpu_library)
+ 
+@@ -1685,7 +1686,7 @@ if(BUILD_SHARED_LIBS)
+       target_link_libraries(torch_global_deps torch::nvtoolsext)
+     endif()
+   endif()
+-  install(TARGETS torch_global_deps DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+  install(TARGETS torch_global_deps RUNTIME DESTINATION "${TORCH_INSTALL_BIN_DIR}" LIBRARY DESTINATION "${TORCH_INSTALL_LIB_DIR}" ARCHIVE DESTINATION "${TORCH_INSTALL_LIB_DIR}")
+ endif()
+ 
+ # ---[ Caffe2 HIP sources.
+diff --git a/cmake/Codegen.cmake b/cmake/Codegen.cmake
+index 724d993..f743939 100644
+--- a/cmake/Codegen.cmake
+++ b/cmake/Codegen.cmake
+@@ -36,11 +36,13 @@ endfunction()
+ ################################################################################
+ 
+ # -- [ Deterine commit hash
+-execute_process(
+-    COMMAND "${Python_EXECUTABLE}" -c "from tools.generate_torch_version import get_sha;print(get_sha('.'), end='')"
+-    OUTPUT_VARIABLE COMMIT_SHA
+-    WORKING_DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/..
+-)
+if(NOT DEFINED COMMIT_SHA) 
+  execute_process(
+      COMMAND "${Python_EXECUTABLE}" -c "from tools.generate_torch_version import get_sha;print(get_sha('.'), end='')"
+      OUTPUT_VARIABLE COMMIT_SHA
+      WORKING_DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/..
+  )
+endif()
+ 
+ # ---[ Write the macros file
+ configure_file(
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 5227204..9acaf75 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -46,6 +46,7 @@ if(USE_CUDA)
+     # A helper variable recording the list of Caffe2 dependent libraries
+     # torch::cudart is dealt with separately, due to CUDA_ADD_LIBRARY
+     # design reason (it adds CUDA_LIBRARIES itself).
+    find_package(NvidiaCutlass CONFIG REQUIRED)
+     set(Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS )
+     if(NOT CAFFE2_USE_NVRTC)
+       caffe2_update_option(USE_NVRTC OFF)
+@@ -378,6 +379,9 @@ if(INTERN_BUILD_MOBILE OR NOT DISABLE_NNPACK_AND_FAMILY)
+   set(USE_PTHREADPOOL ON CACHE BOOL "" FORCE)
+   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_PTHREADPOOL")
+ 
+  find_package(unofficial-pthreadpool CONFIG REQUIRED)
+  add_library(pthreadpool ALIAS unofficial::pthreadpool)
+
+   if(NOT TARGET pthreadpool)
+     if(USE_SYSTEM_PTHREADPOOL)
+       add_library(pthreadpool SHARED IMPORTED)
+@@ -412,6 +416,9 @@ endif()
+ if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "^(s390x|ppc64le)$")
+   # ---[ Caffe2 uses cpuinfo library in the thread pool
+   # ---[ But it doesn't support s390x/powerpc and thus not used on s390x/powerpc
+  find_package(cpuinfo CONFIG REQUIRED)
+  add_library(cpuinfo ALIAS cpuinfo::cpuinfo)
+  add_library(clog ALIAS cpuinfo::clog)
+   if(NOT TARGET cpuinfo AND USE_SYSTEM_CPUINFO)
+     add_library(cpuinfo SHARED IMPORTED)
+     find_library(CPUINFO_LIBRARY cpuinfo)
+@@ -476,8 +483,10 @@ endif()
+ 
+ # ---[ NNPACK
+ if(USE_NNPACK)
+-  include(${CMAKE_CURRENT_LIST_DIR}/External/nnpack.cmake)
+-  if(NNPACK_FOUND)
+
+  find_package(unofficial-nnpack CONFIG REQUIRED)
+  add_library(nnpack ALIAS unofficial::nnpack::nnpack)
+  if(nnpack_FOUND)
+     if(TARGET nnpack)
+       # ---[ NNPACK is being built together with Caffe2: explicitly specify dependency
+       list(APPEND Caffe2_DEPENDENCY_LIBS nnpack)
+@@ -492,6 +501,12 @@ if(USE_NNPACK)
+ endif()
+ 
+ # ---[ XNNPACK
+if(USE_XNNPACK)
+  find_package(unofficial-xnnpack CONFIG REQUIRED)
+  add_library(XNNPACK ALIAS unofficial::xnnpack::XNNPACK)
+  add_library(microkernels-prod ALIAS unofficial::xnnpack::microkernels-prod)
+  list(APPEND Caffe2_DEPENDENCY_LIBS XNNPACK microkernels-prod)
+endif()
+ if(USE_XNNPACK AND NOT USE_SYSTEM_XNNPACK)
+   if(NOT DEFINED XNNPACK_SOURCE_DIR)
+     set(XNNPACK_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/XNNPACK" CACHE STRING "XNNPACK source directory")
+@@ -702,7 +717,7 @@ if(BUILD_TEST OR BUILD_MOBILE_BENCHMARK OR BUILD_MOBILE_TEST)
+ endif()
+ 
+ # ---[ FBGEMM
+-if(USE_FBGEMM)
+if(FALSE)
+   set(CAFFE2_THIRD_PARTY_ROOT "${PROJECT_SOURCE_DIR}/third_party")
+   if(NOT DEFINED FBGEMM_SOURCE_DIR)
+     set(FBGEMM_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/fbgemm" CACHE STRING "FBGEMM source directory")
+@@ -755,6 +770,9 @@ endif()
+ 
+ if(USE_FBGEMM)
+   caffe2_update_option(USE_FBGEMM ON)
+  find_package(fbgemmLibrary REQUIRED)
+  list(APPEND Caffe2_DEPENDENCY_LIBS fbgemm)
+
+ else()
+   caffe2_update_option(USE_FBGEMM OFF)
+   message(WARNING
+@@ -804,6 +822,8 @@ if(USE_ITT)
+ endif()
+ 
+ # ---[ Caffe2 depends on FP16 library for half-precision conversions
+find_package(unofficial-fp16 CONFIG REQUIRED)
+add_library(fp16 ALIAS unofficial::fp16::fp16)
+ if(NOT TARGET fp16 AND NOT USE_SYSTEM_FP16)
+   set(CAFFE2_THIRD_PARTY_ROOT "${PROJECT_SOURCE_DIR}/third_party")
+   # PSIMD is required by FP16
+@@ -1155,6 +1175,7 @@ if(USE_DISTRIBUTED AND USE_TENSORPIPE)
+   if(MSVC)
+     message(WARNING "Tensorpipe cannot be used on Windows.")
+   else()
+    if(FALSE)
+     if(USE_CUDA)
+       set(TP_USE_CUDA ON CACHE BOOL "" FORCE)
+       set(TP_ENABLE_CUDA_IPC ON CACHE BOOL "" FORCE)
+@@ -1177,6 +1198,9 @@ if(USE_DISTRIBUTED AND USE_TENSORPIPE)
+     if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.0.0")
+       unset(CMAKE_POLICY_VERSION_MINIMUM)
+     endif()
+    endif()
+    find_package(unofficial-tensorpipe CONFIG REQUIRED)
+    add_library(tensorpipe ALIAS unofficial::tensorpipe::tensorpipe)
+ 
+     list(APPEND Caffe2_DEPENDENCY_LIBS tensorpipe)
+     list(APPEND Caffe2_DEPENDENCY_LIBS nlohmann)
+@@ -1329,7 +1353,7 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+       caffe2_interface_library(onnx onnx_library)
+     endif()
+     list(APPEND Caffe2_DEPENDENCY_WHOLE_LINK_LIBS onnx_library)
+-  else()
+  elseif(FALSE)
+     add_library(onnx SHARED IMPORTED)
+     find_library(ONNX_LIBRARY onnx)
+     if(NOT ONNX_LIBRARY)
+@@ -1345,6 +1369,8 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT INTERN_DISABLE_ONNX)
+     message("-- Found onnx: ${ONNX_LIBRARY} ${ONNX_PROTO_LIBRARY}")
+     list(APPEND Caffe2_DEPENDENCY_LIBS onnx_proto onnx)
+   endif()
+  find_package(ONNX CONFIG REQUIRED)
+  list(APPEND Caffe2_DEPENDENCY_LIBS ONNX::onnx ONNX::onnx_proto)
+   # Recover the build shared libs option.
+   set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS})
+ endif()
+@@ -1515,6 +1541,9 @@ if(NOT INTERN_BUILD_MOBILE)
+   endif()
+ 
+   if(USE_KLEIDIAI)
+    find_package(KleidiAI CONFIG REQUIRED)
+    list(APPEND Caffe2_DEPENDENCY_LIBS KleidiAI::kleidiai)
+  elseif(FALSE)
+     if(CMAKE_C_COMPILER_ID STREQUAL "Clang" AND CMAKE_C_COMPILER_VERSION VERSION_LESS "11" )
+         message(WARNING "KleidiAI: Using non-supported Clang version. Expected 11 or newer, received ${CMAKE_C_COMPILER_VERSION}.")
+     endif()
+@@ -1586,6 +1615,7 @@ endif()
+ #
+ set(TEMP_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+ set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build shared libs" FORCE)
+if(0)
+ add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+ 
+ # Disable compiler feature checks for `fmt`.
+@@ -1596,7 +1626,8 @@ add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+ # `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 "")
+-
+endif()
+find_package(fmt REQUIRED)
+ list(APPEND Caffe2_DEPENDENCY_LIBS fmt::fmt-header-only)
+ set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
+ 
+@@ -1745,13 +1776,10 @@ if(USE_KINETO)
+   endif()
+ endif()
+ 
+-# Include google/FlatBuffers
+-include(${CMAKE_CURRENT_LIST_DIR}/FlatBuffers.cmake)
+ 
+-# Include cpp-httplib
+-add_library(httplib INTERFACE IMPORTED)
+-target_include_directories(httplib SYSTEM INTERFACE ${PROJECT_SOURCE_DIR}/third_party/cpp-httplib)
+find_package(FlatBuffers REQUIRED)
+add_library(flatbuffers ALIAS flatbuffers::flatbuffers)
+find_package(httplib REQUIRED)
+find_package(nlohmann_json REQUIRED)
+add_library(nlohmann ALIAS nlohmann_json)
+ 
+-# Include nlohmann-json
+-add_library(nlohmann INTERFACE IMPORTED)
+-include_directories(nlohmann SYSTEM INTERFACE ${PROJECT_SOURCE_DIR}/third_party/nlohmann/include)
+diff --git a/torch/CMakeLists.txt b/torch/CMakeLists.txt
+index 8b8ebdc..1777c0e 100644
+--- a/torch/CMakeLists.txt
+++ b/torch/CMakeLists.txt
+@@ -59,18 +59,7 @@ set(TORCH_PYTHON_INCLUDE_DIRECTORIES
+     ${CMAKE_BINARY_DIR}
+     ${CMAKE_BINARY_DIR}/aten/src
+     ${CMAKE_BINARY_DIR}/caffe2/aten/src
+-    ${CMAKE_BINARY_DIR}/third_party
+-    ${CMAKE_BINARY_DIR}/third_party/onnx
+-
+     ${TORCH_ROOT}/third_party/valgrind-headers
+-
+-    ${TORCH_ROOT}/third_party/gloo
+-    ${TORCH_ROOT}/third_party/onnx
+-    ${TORCH_ROOT}/third_party/flatbuffers/include
+-    ${TORCH_ROOT}/third_party/kineto/libkineto/include
+-    ${TORCH_ROOT}/third_party/cpp-httplib
+-    ${TORCH_ROOT}/third_party/nlohmann/include
+-
+     ${TORCH_SRC_DIR}/csrc
+     ${TORCH_SRC_DIR}/csrc/api/include
+     ${TORCH_SRC_DIR}/lib
--- a/ports/libtorch/fix-dist-cuda.patch
+++ b/ports/libtorch/fix-dist-cuda.patch
@ -0,0 +1,22 @@
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 9acaf75..4f44d3e 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1205,6 +1205,8 @@ if(USE_DISTRIBUTED AND USE_TENSORPIPE)
+     list(APPEND Caffe2_DEPENDENCY_LIBS tensorpipe)
+     list(APPEND Caffe2_DEPENDENCY_LIBS nlohmann)
+     if(USE_CUDA)
+      add_library(tensorpipe_cuda ALIAS unofficial::tensorpipe::tensorpipe_cuda)
+
+       list(APPEND Caffe2_CUDA_DEPENDENCY_LIBS tensorpipe_cuda)
+     elseif(USE_ROCM)
+       message(WARNING "TensorPipe doesn't yet support ROCm")
+@@ -1279,7 +1281,7 @@ if(USE_GLOO)
+     endif()
+     # Pick the right dependency depending on USE_CUDA
+     list(APPEND Caffe2_DEPENDENCY_LIBS gloo)
+-    if(USE_CUDA)
+    if(FALSE)
+       list(APPEND Caffe2_CUDA_DEPENDENCY_LIBS gloo_cuda)
+     elseif(USE_ROCM)
+       list(APPEND Caffe2_HIP_DEPENDENCY_LIBS gloo_hip)
--- a/ports/libtorch/fix-glog.patch
+++ b/ports/libtorch/fix-glog.patch
@ -1,8 +1,21 @@
 diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
-index ff8e1d6cc..306dac51f 100644
--- a/c10/util/Logging.cpp	
+index c463658..a523040 100644
+--- a/c10/util/Logging.cpp
 +++ b/c10/util/Logging.cpp
-@@ -218,24 +218,11 @@ C10_DEFINE_int(
+@@ -362,7 +362,7 @@ void UpdateLoggingLevelsFromFlags() {
+ 
+ void ShowLogInfoToStderr() {
+   FLAGS_logtostderr = 1;
+-  FLAGS_minloglevel = std::min(FLAGS_minloglevel, google::GLOG_INFO);
+  FLAGS_minloglevel = std::min<int>(FLAGS_minloglevel, google::GLOG_INFO);
+ }
+ } // namespace c10
+ 
+diff --git a/c10/util/Logging.cpp b/c10/util/Logging.cpp
+index a523040..04375b6 100644
+--- a/c10/util/Logging.cpp
+++ b/c10/util/Logging.cpp
+@@ -295,23 +295,13 @@ C10_DEFINE_int(
     google::GLOG_WARNING,
     "The minimum log level that caffe2 will output.");
 
@ -20,11 +33,10 @@ index ff8e1d6cc..306dac51f 100644
 namespace {
 
 void initGoogleLogging(char const* name) {
-#if !defined(_MSC_VER)
-  // This trick can only be used on UNIX platforms
+ #if !defined(_MSC_VER)
+   // This trick can only be used on UNIX platforms
 -  if (!::google::glog_internal_namespace_::IsGoogleLoggingInitialized())
-#endif
 +  if (!::google::IsGoogleLoggingInitialized())
+ #endif
   {
     ::google::InitGoogleLogging(name);
- #if !defined(_MSC_VER)
--- a/ports/libtorch/fix-msvc-ICE.patch
+++ b/ports/libtorch/fix-msvc-ICE.patch
@ -1,15 +0,0 @@
-diff --git a/aten/src/ATen/cpu/vec/vec_base.h b/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
-@@ -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() {
-    return VECTOR_WIDTH / size_T;
-+    return VECTOR_WIDTH / sizeof(T);
-   }
-   Vectorized() : values{static_cast<T>(0)} {}
-   Vectorized(T val) {
--- a/ports/libtorch/fix-osx.patch
+++ b/ports/libtorch/fix-osx.patch
@ -0,0 +1,16 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 463788e..1ca26e3 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -796,7 +796,10 @@ if(NOT CMAKE_BUILD_TYPE)
+ endif()
+ 
+ # The below means we are cross compiling for arm64 or x86_64 on MacOSX
+-if(NOT IOS
+if(TRUE)
+  message(STATUS "Using custom protoc executable: ${PROTOBUF_PROTOC_EXECUTABLE}")
+  message(STATUS "Using custom caffe2 protoc executable: ${CAFFE2_CUSTOM_PROTOC_EXECUTABLE}")
+elseif(NOT IOS
+    AND CMAKE_SYSTEM_NAME STREQUAL "Darwin"
+    AND CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64|arm64)$")
+   set(CROSS_COMPILING_MACOSX TRUE)
--- a/ports/libtorch/fix-pytorch-pr-156630.patch
+++ b/ports/libtorch/fix-pytorch-pr-156630.patch
@ -0,0 +1,29 @@
+diff --git a/aten/src/ATen/native/cuda/SegmentReduce.cu b/aten/src/ATen/native/cuda/SegmentReduce.cu
+index 04bec04..3ea8af2 100644
+--- a/aten/src/ATen/native/cuda/SegmentReduce.cu
+++ b/aten/src/ATen/native/cuda/SegmentReduce.cu
+@@ -1,5 +1,6 @@
+ #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
+ #include <ATen/native/SegmentReduce.h>
+#include <cuda_runtime.h>
+ 
+ #include <ATen/core/Tensor.h>
+ #include <ATen/Dispatch.h>
+@@ -17,6 +18,10 @@
+ #include <ATen/ops/cumsum.h>
+ #endif
+ 
+// SegmentReduce compilation with CUDA-12.9 causes  NVCC crash on Windows
+// See https://github.com/pytorch/pytorch/issues/156181
+#if !defined(_WIN32) || CUDART_VERSION < 12090
+
+ namespace at::native {
+ 
+ namespace {
+@@ -600,3 +605,5 @@ REGISTER_DISPATCH(
+   &_segment_reduce_offsets_backward_cuda_kernel);
+ 
+ } // namespace at::native
+
+#endif // !defined(_WIN32) || CUDART_VERSION < 12090
+\ No newline at end of file
--- a/ports/libtorch/fix-sleef.patch
+++ b/ports/libtorch/fix-sleef.patch
@ -1,84 +0,0 @@
-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})
-diff --git a/aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h b/aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h
-index f3281d3cb..c8b6ed30e 100644
--- a/aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h
-+++ b/aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h
-@@ -21,6 +21,17 @@ inline namespace CPU_CAPABILITY {
- 
- #if defined(CPU_CAPABILITY_AVX2) && !defined(_MSC_VER)
- 
-+#ifndef SLEEF_CONST
-+#if (defined(__GNUC__) || defined(__CLANG__)) && !defined(__INTEL_COMPILER)
-+#define SLEEF_CONST const
-+#else
-+#define SLEEF_CONST
-+#endif
-+#define SLEEF_CONST_OLD SLEEF_CONST
-+#else
-+#define SLEEF_CONST_OLD
-+#endif
-+
- // bfloat16 conversion
- static inline void cvtbf16_fp32(const __m128i& a, __m256& o) {
-   o = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(a), 16));
-@@ -266,7 +277,7 @@ public:
-     }
-     return b;
-   }
-  Vectorized<T> map(const __m256 (*const vop)(__m256)) const {
-+  Vectorized<T> map(SLEEF_CONST __m256 (*SLEEF_CONST_OLD vop)(__m256)) const {
-     __m256 lo, hi;
-     cvt_to_fp32<T>(values, lo, hi);
-     const auto o1 = vop(lo);
-     diff --git a/aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h b/aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h
-index f3281d3cb..c8b6ed30e 100644
--- a/aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h
-+++ b/aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h
-@@ -18,6 +18,17 @@ inline namespace CPU_CAPABILITY {
- 
- #if defined(CPU_CAPABILITY_AVX512) && !defined(_MSC_VER)
- 
-+#ifndef SLEEF_CONST
-+#if (defined(__GNUC__) || defined(__CLANG__)) && !defined(__INTEL_COMPILER)
-+#define SLEEF_CONST const
-+#else
-+#define SLEEF_CONST
-+#endif
-+#define SLEEF_CONST_OLD SLEEF_CONST
-+#else
-+#define SLEEF_CONST_OLD
-+#endif
-+
- // bfloat16 conversion
- static inline void cvtbf16_fp32(const __m256i& a, __m512& o) {
-   o = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(a), 16));
-@@ -345,7 +356,7 @@ public:
-   }
-   #pragma clang diagnostic push
-   #pragma clang diagnostic ignored "-Wignored-qualifiers"
-  Vectorized<T> map(const __m512 (*const vop)(__m512)) const {
-+  Vectorized<T> map(SLEEF_CONST __m512 (*SLEEF_CONST_OLD vop)(__m512)) const {
-     __m512 lo, hi;
-     cvt_to_fp32<T>(values, lo, hi);
-     const auto o1 = vop(lo);
--- a/ports/libtorch/fix-vulkan.patch
+++ b/ports/libtorch/fix-vulkan.patch
@ -0,0 +1,43 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 1ca26e3..7d3442c 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -1104,7 +1104,8 @@ if(NOT MSVC)
+ else()
+   # Define export functions for AOTI.
+   add_compile_definitions(EXPORT_AOTI_FUNCTIONS)
+-
+  # needed for vulkan compile
+  add_compile_definitions(_USE_MATH_DEFINES) # math.h macros like M_PI, M_SQRT2, M_2_SQRTPI, etc.
+   # skip unwanted includes from windows.h
+   add_compile_definitions(WIN32_LEAN_AND_MEAN)
+   # Windows SDK broke compatibility since version 25131, but introduced this
+diff --git a/aten/src/ATen/native/vulkan/api/Allocator.h b/aten/src/ATen/native/vulkan/api/Allocator.h
+index a89c333..a9a41ef 100644
+--- a/aten/src/ATen/native/vulkan/api/Allocator.h
+++ b/aten/src/ATen/native/vulkan/api/Allocator.h
+@@ -50,7 +50,11 @@
+ #pragma clang diagnostic ignored "-Winconsistent-missing-destructor-override"
+ #endif /* __clang__ */
+ 
+-#include <include/vk_mem_alloc.h>
+#if __has_include(<vma/vk_mem_alloc.h>)
+#include <vma/vk_mem_alloc.h>
+#else
+#include <vk_mem_alloc.h>
+#endif
+ 
+ #ifdef __clang__
+ #pragma clang diagnostic pop
+diff --git a/aten/src/ATen/native/vulkan/api/QueryPool.cpp b/aten/src/ATen/native/vulkan/api/QueryPool.cpp
+index 9c0c7fb..ae4a81f 100644
+--- a/aten/src/ATen/native/vulkan/api/QueryPool.cpp
+++ b/aten/src/ATen/native/vulkan/api/QueryPool.cpp
+@@ -9,6 +9,7 @@
+ #include <iomanip>
+ #include <iostream>
+ #include <utility>
+#include <algorithm>
+ 
+ namespace at {
+ namespace native {
--- a/ports/libtorch/force-cuda-include.patch
+++ b/ports/libtorch/force-cuda-include.patch
@ -1,12 +0,0 @@
-diff --git a/cmake/public/cuda.cmake b/cmake/public/cuda.cmake
-index 5964429..d528597 100644
--- a/cmake/public/cuda.cmake
-+++ b/cmake/public/cuda.cmake
-@@ -57,6 +57,7 @@ find_package(CUDAToolkit REQUIRED)
- 
- cmake_policy(POP)
- 
-+set(CUDAToolkit_INCLUDE_DIR "${CUDA_INCLUDE_DIRS}")
- if(NOT CMAKE_CUDA_COMPILER_VERSION STREQUAL CUDAToolkit_VERSION OR
-     NOT CUDA_INCLUDE_DIRS STREQUAL CUDAToolkit_INCLUDE_DIR)
-   message(FATAL_ERROR "Found two conflicting CUDA installs:\n"
--- a/ports/libtorch/fxdiv.patch
+++ b/ports/libtorch/fxdiv.patch
@ -1,45 +0,0 @@
-diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
-index d185233..11f285b 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 f7a595e..47faec5 100644
--- a/caffe2/CMakeLists.txt
-+++ b/caffe2/CMakeLists.txt
-@@ -108,12 +108,19 @@ 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/more-fixes.patch
+++ b/ports/libtorch/more-fixes.patch
@ -1,320 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 7b8bc7c..fa6da32 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 d0c8bdf..50df97a 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 dc22b98..f91fbf6 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 ad407e1..bbca41b 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 74d0d55..cf349c2 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/Modules/FindCUSPARSELT.cmake b/cmake/Modules/FindCUSPARSELT.cmake
-index 8727f44..1e41281 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 d2bff53..3af9012 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 2376f1b..a8ae436 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 ${JIT_TEST_DEPENDENCIES})
-diff --git a/third_party/nvfuser/CMakeLists.txt b/third_party/nvfuser/CMakeLists.txt
-index b148418..7b96083 100644
--- a/third_party/nvfuser/CMakeLists.txt
-+++ b/third_party/nvfuser/CMakeLists.txt
-@@ -158,7 +158,10 @@ endif()
- target_include_directories(${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 62ee4c1..f75073c 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 4fb72c5..1abe1b7 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 7ab506f..2ee3b68 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 dec8757..3bb368a 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/no-abs-path.patch
+++ b/ports/libtorch/no-abs-path.patch
@ -1,25 +0,0 @@
-diff --git a/caffe2/core/macros.h.in b/caffe2/core/macros.h.in
-index 997752ede3..7209b9a63e 100644
--- a/caffe2/core/macros.h.in	
-+++ b/caffe2/core/macros.h.in
-@@ -35,7 +35,7 @@
- // Useful build settings that are recorded in the compiled binary
- #define CAFFE2_BUILD_STRINGS { \
-   {"TORCH_VERSION", "${TORCH_VERSION}"}, \
-  {"CXX_COMPILER", "${CMAKE_CXX_COMPILER}"}, \
-+  {"CXX_COMPILER", "${cxx_compiler}"}, \
-   {"CXX_FLAGS", "${CMAKE_CXX_FLAGS}"}, \
-   {"BUILD_TYPE", "${CMAKE_BUILD_TYPE}"}, \
-   {"BLAS_INFO", "${BLAS_INFO}"}, \
-diff --git a/cmake/Codegen.cmake b/cmake/Codegen.cmake
-index 6fa2ab6ea1..07c449c02b 100644
--- a/cmake/Codegen.cmake	
-+++ b/cmake/Codegen.cmake
-@@ -36,6 +36,7 @@ endfunction()
- ################################################################################
- 
- # ---[ Write the macros file
-+cmake_path(GET CMAKE_CXX_COMPILER FILENAME cxx_compiler)
- configure_file(
-     ${CMAKE_CURRENT_LIST_DIR}/../caffe2/core/macros.h.in
-     ${CMAKE_BINARY_DIR}/caffe2/core/macros.h)
--- a/ports/libtorch/portfile.cmake
+++ b/ports/libtorch/portfile.cmake
@ -1,285 +1,240 @@
-vcpkg_download_distfile(
-    CUDNN_9_FIX
-    URLS https://github.com/pytorch/pytorch/commit/e14026bc2a6cd80bedffead77a5d7b75a37f8e67.patch?full_index=1
-    SHA512 9569547b44b61f9559f0e7ab91f2be51657ece4f5462b6860cb5eae8d23d01187d6af046b369a77a228fe4d7153f5c683b686e84c1296a662f83e5f1f281bc7e
-    FILENAME libtorch-cudnn-9-fix-e14026bc2a6cd80bedffead77a5d7b75a37f8e67.patch
-)
-
-vcpkg_download_distfile(
-    CUDA_THRUST_MISSING_HEADER_FIX
-    URLS https://github.com/pytorch/pytorch/commit/2a440348958b3f0a2b09458bd76fe5959b371c0c.patch?full_index=1
-    SHA512 eff10d81b1c635108ad1b95a430865a76ab3f2079be74e61e06876942ac1fd43a274fc1c73e43c2c01b9ce5aca648213ef75c13c28b8ffa40497e4e26d5e3b16
-    FILENAME libtorch-cuda-thrust-missing-header-2a440348958b3f0a2b09458bd76fe5959b371c0c.patch
-)
-
-vcpkg_check_linkage(ONLY_DYNAMIC_LIBRARY)
-
-vcpkg_from_github(
-    OUT_SOURCE_PATH SOURCE_PATH
-    REPO pytorch/pytorch
-    REF "v${VERSION}"
-    SHA512 a8961d78ad785b13c959a0612563a60e0de17a7c8bb9822ddea9a24072796354d07e81c47b6cc8761b21a6448845b088cf80e1661d9e889b0ed5474d3dc76756
-    HEAD_REF master
-    PATCHES
-        "${CUDNN_9_FIX}"
-        "${CUDA_THRUST_MISSING_HEADER_FIX}"
-        cmake-fixes.patch
-        more-fixes.patch
-        fix-build.patch
-        clang-cl.patch
-        cuda-adjustments.patch
-        fix-api-export.patch
-        fxdiv.patch
-        protoc.patch
-        fix-sleef.patch
-        fix-glog.patch
-        fix-msvc-ICE.patch
-        fix-calculate-minloglevel.patch
-        force-cuda-include.patch
-        fix-aten-cutlass.patch
-        fix-build-error-with-fmt11.patch
-        no-abs-path.patch
-        add-include-chrono.patch
-)
-
-file(REMOVE_RECURSE "${SOURCE_PATH}/caffe2/core/macros.h") # We must use generated header files
-
-vcpkg_from_github(
-    OUT_SOURCE_PATH src_kineto
-    REPO pytorch/kineto
-    REF 49e854d805d916b2031e337763928d2f8d2e1fbf
-    SHA512 ae63d48dc5b8ac30c38c2ace60f16834c7e9275fa342dc9f109d4fbc87b7bd674664f6413c36d0c1ab5a7da786030a4108d83daa4502b2f30239283ea3acdb16
-    HEAD_REF main
-    PATCHES
-      kineto.patch
-)
-file(COPY "${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 "${SOURCE_PATH}/third_party/cudnn_frontend")
-
-
-file(REMOVE
-  "${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 ${FLATC} --cpp --no-prefix --scoped-enums --gen-mutable mobile_bytecode.fbs
-    LOGNAME codegen-flatc-mobile_bytecode
-    WORKING_DIRECTORY "${SOURCE_PATH}/torch/csrc/jit/serialization"
-)
-
-find_program(PROTOC NAMES protoc PATHS "${CURRENT_HOST_INSTALLED_DIR}/tools/protobuf" REQUIRED NO_DEFAULT_PATH NO_CMAKE_PATH)
-message(STATUS "Using protoc: ${PROTOC}")
-
-x_vcpkg_get_python_packages(
-    PYTHON_VERSION 3
-    PACKAGES typing-extensions pyyaml numpy
-    OUT_PYTHON_VAR PYTHON3
-)
-#set(PYTHON3 "${CURRENT_HOST_INSTALLED_DIR}/tools/python3/python${VCPKG_HOST_EXECUTABLE_SUFFIX}")
-message(STATUS "Using Python3: ${PYTHON3}")
-
-vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS
-  FEATURES
-    dist    USE_DISTRIBUTED # MPI, Gloo, TensorPipe
-    zstd    USE_ZSTD
-    fbgemm  USE_FBGEMM
-    opencv  USE_OPENCV
-    # These are alternatives !
-    # tbb     USE_TBB
-    # tbb     AT_PARALLEL_NATIVE_TBB # AT_PARALLEL_ are alternatives
-    # openmp  USE_OPENMP
-    # openmp  AT_PARALLEL_OPENMP # AT_PARALLEL_ are alternatives
-    leveldb USE_LEVELDB
-    opencl  USE_OPENCL
-    cuda    USE_CUDA
-    cuda    USE_CUDNN
-    cuda    USE_NCCL
-    cuda    USE_SYSTEM_NCCL
-    cuda    USE_NVRTC
-    cuda    AT_CUDA_ENABLED
-    cuda    AT_CUDNN_ENABLED
-    cuda    USE_MAGMA
-    vulkan  USE_VULKAN
-    #vulkan  USE_VULKAN_SHADERC_RUNTIME
-    vulkan  USE_VULKAN_RELAXED_PRECISION
-    rocm    USE_ROCM  # This is an alternative to cuda not a feature! (Not in vcpkg.json!) -> disabled
-    llvm    USE_LLVM
-    mpi     USE_MPI
-    nnpack  USE_NNPACK  # todo: check use of `DISABLE_NNPACK_AND_FAMILY`
-    nnpack  AT_NNPACK_ENABLED
-    qnnpack USE_QNNPACK # todo: check use of `USE_PYTORCH_QNNPACK`
-#   No feature in vcpkg yet so disabled. -> Requires numpy build by vcpkg itself
-    python  BUILD_PYTHON
-    python  USE_NUMPY
-)
-
-if("dist" IN_LIST FEATURES)
-    if(VCPKG_TARGET_IS_LINUX OR VCPKG_TARGET_IS_OSX)
-        list(APPEND FEATURE_OPTIONS -DUSE_TENSORPIPE=ON)
-    endif()
-    if(VCPKG_TARGET_IS_WINDOWS OR VCPKG_TARGET_IS_OSX)
-        list(APPEND FEATURE_OPTIONS -DUSE_LIBUV=ON)
-    endif()
-    list(APPEND FEATURE_OPTIONS -DUSE_GLOO=${VCPKG_TARGET_IS_LINUX})
-endif()
-
-if(VCPKG_TARGET_IS_ANDROID OR VCPKG_TARGET_IS_IOS)
-    list(APPEND FEATURE_OPTIONS -DINTERN_BUILD_MOBILE=ON)
-else()
-    list(APPEND FEATURE_OPTIONS -DINTERN_BUILD_MOBILE=OFF)
-endif()
-
-string(COMPARE EQUAL "${VCPKG_CRT_LINKAGE}" "static" USE_STATIC_RUNTIME)
-
-vcpkg_cmake_configure(
-    SOURCE_PATH "${SOURCE_PATH}"
-    DISABLE_PARALLEL_CONFIGURE
-    OPTIONS
-        ${FEATURE_OPTIONS}
-        -DProtobuf_PROTOC_EXECUTABLE:FILEPATH=${PROTOC}
-        -DCAFFE2_CUSTOM_PROTOC_EXECUTABLE:FILEPATH=${PROTOC}
-        -DPYTHON_EXECUTABLE:FILEPATH=${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
-        -DUSE_METAL=OFF
-        -DUSE_PYTORCH_METAL=OFF
-        -DUSE_PYTORCH_METAL_EXPORT=OFF
-        -DUSE_GFLAGS=ON
-        -DUSE_GLOG=ON
-        -DUSE_LMDB=ON
-        -DUSE_ITT=OFF
-        -DUSE_ROCKSDB=ON
-        -DUSE_OBSERVERS=OFF
-        -DUSE_PYTORCH_QNNPACK=OFF
-        -DUSE_KINETO=OFF
-        -DUSE_ROCM=OFF
-        -DUSE_NUMA=OFF
-        -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_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}
-        ${BLAS_OPTIONS}
-        # BLAS=MKL not supported in this port
-        -DUSE_MKLDNN=OFF
-        -DUSE_MKLDNN_CBLAS=OFF
-        #-DCAFFE2_USE_MKL=ON
-        #-DAT_MKL_ENABLED=ON
-        -DAT_MKLDNN_ENABLED=OFF
-        -DUSE_OPENCL=ON
-        -DUSE_NUMPY=ON
-        -DUSE_KINETO=OFF #
-    OPTIONS_RELEASE
-      -DPYTHON_LIBRARY=${CURRENT_INSTALLED_DIR}/lib/python311.lib
-    OPTIONS_DEBUG
-      -DPYTHON_LIBRARY=${CURRENT_INSTALLED_DIR}/debug/lib/python311_d.lib
-    MAYBE_UNUSED_VARIABLES
-        USE_NUMA
-        USE_SYSTEM_BIND11
-        MKLDNN_CPU_RUNTIME
-        PYTHON_LIBRARY
-)
-
-vcpkg_cmake_install()
-vcpkg_copy_pdbs()
-
-vcpkg_cmake_config_fixup(PACKAGE_NAME caffe2 CONFIG_PATH "share/cmake/Caffe2" DO_NOT_DELETE_PARENT_CONFIG_PATH)
-vcpkg_cmake_config_fixup(PACKAGE_NAME torch CONFIG_PATH "share/cmake/Torch")
-vcpkg_replace_string("${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 "${folder}")
-        return()
-    endif()
-    file(GLOB paths LIST_DIRECTORIES true "${folder}/*")
-    list(LENGTH paths count)
-    # 1. remove if the given folder is empty
-    if(count EQUAL 0)
-        file(REMOVE_RECURSE "${folder}")
-        message(STATUS "Removed ${folder}")
-        return()
-    endif()
-    # 2. repeat the operation for hop 1 sub-directories 
-    foreach(path ${paths})
-        cleanup_once(${path})
-    endforeach()
-endfunction()
-
-# Some folders may contain empty folders. They will become empty after `cleanup_once`.
-# Repeat given times to delete new empty folders.
-function(cleanup_repeat folder repeat)
-    if(NOT IS_DIRECTORY "${folder}")
-        return()
-    endif()
-    while(repeat GREATER_EQUAL 1)
-        math(EXPR repeat "${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"
-    "${CURRENT_PACKAGES_DIR}/debug/share"
-)
-
-vcpkg_install_copyright(FILE_LIST "${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("${CURRENT_PACKAGES_DIR}/include/torch/csrc/autograd/custom_function.h" "struct TORCH_API Function" "struct Function")
+vcpkg_check_linkage(ONLY_DYNAMIC_LIBRARY)
+
+vcpkg_from_github(
+    OUT_SOURCE_PATH SOURCE_PATH
+    REPO pytorch/pytorch
+    REF "v${VERSION}"
+    SHA512 a9fc2252af9031c2cd46dde558c491aea8bc322fb80157a7760f300a44b759d4bfe866f030fbb974b80493057cfff4dd512498f99a100ed6d05bf620258ed37e
+    HEAD_REF master
+    PATCHES
+        fix-cmake.patch
+        fix-osx.patch
+        fix-vulkan.patch
+        fix-glog.patch
+        fix-pytorch-pr-156630.patch # https://github.com/pytorch/pytorch/pull/156630
+        fix-dist-cuda.patch
+        )
+
+file(REMOVE_RECURSE "${SOURCE_PATH}/caffe2/core/macros.h") # We must use generated header files
+
+vcpkg_from_github(
+    OUT_SOURCE_PATH src_kineto
+    REPO pytorch/kineto
+    REF d9753139d181b9ff42872465aac0e5d3018be415
+    SHA512 f037fac78e566c40108acf9eace55a8f67a2c5b71f298fd3cd17bf22cf05240c260fd89f017fa411656a7505ec9073a06a3048e191251d5cfc4b52c237b37d0b
+    HEAD_REF main
+    PATCHES
+      kineto.patch
+)
+file(COPY "${src_kineto}/" DESTINATION "${SOURCE_PATH}/third_party/kineto")
+
+vcpkg_from_github(
+    OUT_SOURCE_PATH src_cudnn
+    REPO NVIDIA/cudnn-frontend # new port ?
+    REF 2533f5e5c1877fd76266133c1479ef1643ce3a8b #  1.6.1 
+    SHA512 8caacdf9f7dbd6ce55507f5f7165db8640b681e2a7dfd6a841de8eaa3489cff5ba41d11758cc464320b2ff9a491f8234e1749580cf43cac702f07cf82611e084
+    HEAD_REF main
+)
+file(COPY "${src_cudnn}/" DESTINATION "${SOURCE_PATH}/third_party/cudnn_frontend")
+
+
+file(REMOVE
+  "${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 ${FLATC} --cpp --no-prefix --scoped-enums --gen-mutable mobile_bytecode.fbs
+    LOGNAME codegen-flatc-mobile_bytecode
+    WORKING_DIRECTORY "${SOURCE_PATH}/torch/csrc/jit/serialization"
+)
+
+find_program(PROTOC NAMES protoc PATHS "${CURRENT_HOST_INSTALLED_DIR}/tools/protobuf" REQUIRED NO_DEFAULT_PATH NO_CMAKE_PATH)
+message(STATUS "Using protoc: ${PROTOC}")
+
+x_vcpkg_get_python_packages(
+    PYTHON_VERSION 3
+    PACKAGES typing-extensions pyyaml 
+    # numpy
+    OUT_PYTHON_VAR PYTHON3
+)
+
+message(STATUS "Using Python3: ${PYTHON3}")
+
+vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS
+  FEATURES
+    dist    USE_DISTRIBUTED # MPI, Gloo, TensorPipe
+    zstd    USE_ZSTD
+    fbgemm  USE_FBGEMM
+    opencv  USE_OPENCV
+    opencl  USE_OPENCL
+    mkldnn  USE_MKLDNN
+    cuda    USE_CUDA
+    cuda    USE_CUDNN
+    cuda    USE_NCCL
+    cuda    USE_SYSTEM_NCCL
+    cuda    USE_NVRTC
+    cuda    AT_CUDA_ENABLED
+    cuda    AT_CUDNN_ENABLED
+    cuda    USE_MAGMA
+    vulkan  USE_VULKAN
+    vulkan  USE_VULKAN_RELAXED_PRECISION
+    rocm    USE_ROCM  # This is an alternative to cuda not a feature! (Not in vcpkg.json!) -> disabled
+    llvm    USE_LLVM
+    mpi     USE_MPI
+    nnpack  USE_NNPACK  # todo: check use of `DISABLE_NNPACK_AND_FAMILY`
+#   No feature in vcpkg yet so disabled. -> Requires numpy build by vcpkg itself
+    python  BUILD_PYTHON
+    python  USE_NUMPY
+    glog    USE_GLOG
+    gflags  USE_GFLAGS
+)
+
+if("dist" IN_LIST FEATURES)
+    if(VCPKG_TARGET_IS_LINUX OR VCPKG_TARGET_IS_OSX)
+        list(APPEND FEATURE_OPTIONS -DUSE_TENSORPIPE=ON)
+    endif()
+    if(VCPKG_TARGET_IS_OSX)
+        list(APPEND FEATURE_OPTIONS -DUSE_LIBUV=ON)
+    endif()
+    list(APPEND FEATURE_OPTIONS -DUSE_GLOO=${VCPKG_TARGET_IS_LINUX})
+endif()
+
+if("cuda" IN_LIST FEATURES)
+  vcpkg_find_cuda(OUT_CUDA_TOOLKIT_ROOT cuda_toolkit_root)
+    list(APPEND FEATURE_OPTIONS
+        "-DCMAKE_CUDA_COMPILER=${NVCC}"
+        "-DCUDAToolkit_ROOT=${cuda_toolkit_root}"
+    )
+endif()
+
+if("vulkan" IN_LIST FEATURES) # Vulkan::glslc in FindVulkan.cmake
+    find_program(GLSLC NAMES glslc PATHS "${CURRENT_HOST_INSTALLED_DIR}/tools/shaderc" REQUIRED)
+    message(STATUS "Using glslc: ${GLSLC}")
+    list(APPEND FEATURE_OPTIONS "-DVulkan_GLSLC_EXECUTABLE:FILEPATH=${GLSLC}")
+endif()
+
+set(TARGET_IS_MOBILE OFF)
+if(VCPKG_TARGET_IS_ANDROID OR VCPKG_TARGET_IS_IOS)
+    set(TARGET_IS_MOBILE ON)
+endif()
+
+set(TARGET_IS_APPLE OFF)
+if(VCPKG_TARGET_IS_IOS OR VCPKG_TARGET_IS_OSX)
+    set(TARGET_IS_APPLE ON)
+endif()
+
+string(COMPARE EQUAL "${VCPKG_CRT_LINKAGE}" "static" USE_STATIC_RUNTIME)
+vcpkg_cmake_configure(
+    SOURCE_PATH "${SOURCE_PATH}"
+    DISABLE_PARALLEL_CONFIGURE
+    OPTIONS
+        ${FEATURE_OPTIONS}
+        -DProtobuf_PROTOC_EXECUTABLE:FILEPATH=${PROTOC}
+        -DCAFFE2_CUSTOM_PROTOC_EXECUTABLE:FILEPATH=${PROTOC}
+        -DPython_EXECUTABLE:FILEPATH=${PYTHON3}
+        -DPython3_EXECUTABLE:FILEPATH=${PYTHON3}
+        -DBUILD_PYTHON=OFF
+        -DUSE_NUMPY=OFF
+        -DCAFFE2_STATIC_LINK_CUDA=ON
+        -DCAFFE2_USE_MSVC_STATIC_RUNTIME=${USE_STATIC_RUNTIME}
+        -DBUILD_CUSTOM_PROTOBUF=OFF
+        -DBUILD_PYTHON=OFF
+        -DUSE_LITE_PROTO=OFF
+        -DBUILD_TEST=OFF
+        -DATEN_NO_TEST=ON
+        -DUSE_SYSTEM_LIBS=ON
+        -DUSE_METAL=OFF
+        -DUSE_FLASH_ATTENTION=OFF
+        -DUSE_PYTORCH_METAL=OFF
+        -DUSE_PYTORCH_METAL_EXPORT=OFF
+        -DUSE_PYTORCH_QNNPACK:BOOL=OFF
+        -DUSE_ITT=OFF
+        -DUSE_ROCKSDB=ON
+        -DUSE_OBSERVERS=OFF
+        -DUSE_KINETO=OFF
+        -DUSE_ROCM=OFF
+        -DUSE_NUMA=OFF
+        -DUSE_SYSTEM_LIBS=ON
+        -DBUILD_JNI=${VCPKG_TARGET_IS_ANDROID}
+        -DUSE_NNAPI=${VCPKG_TARGET_IS_ANDROID}
+        ${BLAS_OPTIONS}
+        # BLAS=MKL not supported in this port
+        -DUSE_MKLDNN=OFF
+        -DUSE_MKLDNN_CBLAS=OFF
+        #-DCAFFE2_USE_MKL=ON
+        #-DAT_MKL_ENABLED=ON
+        -DAT_MKLDNN_ENABLED=OFF
+        -DUSE_OPENCL=ON
+        -DUSE_KINETO=OFF #
+    # Should be enabled in-future along with the "python" feature (currently disabled)
+    # OPTIONS_RELEASE
+    #  -DPYTHON_LIBRARY=${CURRENT_INSTALLED_DIR}/lib/python311.lib
+    # OPTIONS_DEBUG
+    #  -DPYTHON_LIBRARY=${CURRENT_INSTALLED_DIR}/debug/lib/python311_d.lib
+    MAYBE_UNUSED_VARIABLES
+        USE_NUMA
+        USE_SYSTEM_BIND11
+        MKLDNN_CPU_RUNTIME
+        PYTHON_LIBRARY
+)
+
+vcpkg_cmake_install()
+vcpkg_copy_pdbs()
+
+vcpkg_cmake_config_fixup(PACKAGE_NAME Caffe2 CONFIG_PATH "share/cmake/Caffe2" DO_NOT_DELETE_PARENT_CONFIG_PATH)
+vcpkg_cmake_config_fixup(PACKAGE_NAME torch CONFIG_PATH "share/cmake/Torch" DO_NOT_DELETE_PARENT_CONFIG_PATH)
+vcpkg_cmake_config_fixup(PACKAGE_NAME ATen CONFIG_PATH "share/cmake/ATen" )
+
+vcpkg_replace_string("${CURRENT_PACKAGES_DIR}/share/torch/TorchConfig.cmake" "/../../../" "/../../")
+
+# Traverse the folder and remove "some" empty folders
+function(cleanup_once folder)
+    if(NOT IS_DIRECTORY "${folder}")
+        return()
+    endif()
+    file(GLOB paths LIST_DIRECTORIES true "${folder}/*")
+    list(LENGTH paths count)
+    # 1. remove if the given folder is empty
+    if(count EQUAL 0)
+        file(REMOVE_RECURSE "${folder}")
+        message(STATUS "Removed ${folder}")
+        return()
+    endif()
+    # 2. repeat the operation for hop 1 sub-directories 
+    foreach(path ${paths})
+        cleanup_once(${path})
+    endforeach()
+endfunction()
+
+# Some folders may contain empty folders. They will become empty after `cleanup_once`.
+# Repeat given times to delete new empty folders.
+function(cleanup_repeat folder repeat)
+    if(NOT IS_DIRECTORY "${folder}")
+        return()
+    endif()
+    while(repeat GREATER_EQUAL 1)
+        math(EXPR repeat "${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"
+    "${CURRENT_PACKAGES_DIR}/debug/share"
+)
+
+vcpkg_install_copyright(FILE_LIST "${SOURCE_PATH}/LICENSE")
+
+
+set(VCPKG_POLICY_DLLS_WITHOUT_EXPORTS enabled) # torch_global_deps.dll is empty.c and just for linking deps
+
--- a/ports/libtorch/protoc.patch
+++ b/ports/libtorch/protoc.patch
@ -1,13 +0,0 @@
-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/vcpkg.json
+++ b/ports/libtorch/vcpkg.json
@ -1,16 +1,19 @@
 {
  "name": "libtorch",
-  "version": "2.1.2",
-  "port-version": 12,
+  "version": "2.7.1",
  "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",
+    "cpp-httplib",
    "cpuinfo",
    "eigen3",
-    "fftw3",
+    {
+      "name": "fbgemm",
+      "platform": "x64"
+    },
    "flatbuffers",
    {
      "name": "flatbuffers",
@ -18,14 +21,12 @@
    },
    "fmt",
    "foxi",
-    "foxi",
    "fp16",
    "gemmlowp",
-    "gflags",
-    "glog",
    "lapack",
    "lmdb",
    "mimalloc",
+    "nlohmann-json",
    "onnx",
    "onnx-optimizer",
    "opencl",
@ -36,6 +37,7 @@
      "host": true
    },
    "pthreadpool",
+    "pybind11",
    "sleef",
    {
      "name": "vcpkg-cmake",
@ -52,21 +54,43 @@
    "xnnpack"
  ],
  "default-features": [
+    "gflags",
+    "glog",
    "opencv",
    "zstd"
  ],
  "features": {
    "cuda": {
      "description": "Build with CUDA GPU backend",
+      "supports": "(windows & x64 & !staticcrt) | (linux & x64)",
      "dependencies": [
        "cuda",
        "cudnn",
+        {
+          "name": "gloo",
+          "features": [
+            "cuda"
+          ],
+          "platform": "linux"
+        },
        "magma",
-        "nvidia-cutlass"
+        {
+          "name": "nccl",
+          "platform": "linux"
+        },
+        "nvidia-cutlass",
+        {
+          "name": "tensorpipe",
+          "features": [
+            "cuda"
+          ],
+          "platform": "linux"
+        }
      ]
    },
    "dist": {
      "description": "Use distributed with MPI, Gloo, libuv, TensorPipe",
+      "supports": "linux | windows",
      "dependencies": [
        {
          "name": "gloo",
@ -86,15 +110,20 @@
        },
        {
          "name": "tensorpipe",
-          "platform": "linux | osx"
+          "platform": "linux"
        }
      ]
    },
-    "leveldb": {
-      "description": "Build with LevelDB",
+    "gflags": {
+      "description": "Build with gflags",
      "dependencies": [
-        "leveldb",
-        "snappy"
+        "gflags"
+      ]
+    },
+    "glog": {
+      "description": "Build with glog",
+      "dependencies": [
+        "glog"
      ]
    },
    "llvm": {
@ -125,6 +154,10 @@
    "vulkan": {
      "description": "Build with Vulkan GPU backend",
      "dependencies": [
+        {
+          "name": "shaderc",
+          "host": true
+        },
        "vulkan",
        "vulkan-loader",
        "vulkan-memory-allocator"
--- a/ports/qnnpack/use-packages.patch
+++ b/ports/qnnpack/use-packages.patch
@ -1,5 +1,5 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index a5ddc49..05e1279 100644
+index 932f9e7..187be72 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
@@ -13,8 +13,6 @@ PROJECT(QNNPACK C CXX ASM)
@ -11,12 +11,7 @@ index a5ddc49..05e1279 100644
 OPTION(QNNPACK_BUILD_TESTS "Build QNNPACK unit tests" ON)
 OPTION(QNNPACK_BUILD_BENCHMARKS "Build QNNPACK benchmarks" ON)
 
-@@ -53,11 +51,11 @@ SET(CONFU_DEPENDENCIES_SOURCE_DIR "${CMAKE_SOURCE_DIR}/deps"
- SET(CONFU_DEPENDENCIES_BINARY_DIR "${CMAKE_BINARY_DIR}/deps"
-   CACHE PATH "Confu-style dependencies binary directory")
- 
-IF(NOT DEFINED CLOG_SOURCE_DIR)
-+IF(FALSE)
+@@ -60,7 +58,7 @@ IF(NOT DEFINED CLOG_SOURCE_DIR)
   SET(CLOG_SOURCE_DIR "${PROJECT_SOURCE_DIR}/deps/clog")
 ENDIF()
 
@ -25,7 +20,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading cpuinfo to ${CONFU_DEPENDENCIES_SOURCE_DIR}/cpuinfo (define CPUINFO_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadCpuinfo.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/cpuinfo-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -67,7 +65,7 @@ IF(NOT DEFINED CPUINFO_SOURCE_DIR)
+@@ -70,7 +68,7 @@ IF(NOT DEFINED CPUINFO_SOURCE_DIR)
   SET(CPUINFO_SOURCE_DIR "${CONFU_DEPENDENCIES_SOURCE_DIR}/cpuinfo" CACHE STRING "cpuinfo source directory")
 ENDIF()
 
@ -34,7 +29,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading FP16 to ${CONFU_DEPENDENCIES_SOURCE_DIR}/fp16 (define FP16_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadFP16.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/fp16-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -77,7 +75,7 @@ IF(NOT DEFINED FP16_SOURCE_DIR)
+@@ -80,7 +78,7 @@ IF(NOT DEFINED FP16_SOURCE_DIR)
   SET(FP16_SOURCE_DIR "${CONFU_DEPENDENCIES_SOURCE_DIR}/fp16" CACHE STRING "FP16 source directory")
 ENDIF()
 
@ -43,7 +38,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading FXdiv to ${CONFU_DEPENDENCIES_SOURCE_DIR}/fxdiv (define FXDIV_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadFXdiv.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/fxdiv-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -87,7 +85,7 @@ IF(NOT DEFINED FXDIV_SOURCE_DIR)
+@@ -90,7 +88,7 @@ IF(NOT DEFINED FXDIV_SOURCE_DIR)
   SET(FXDIV_SOURCE_DIR "${CONFU_DEPENDENCIES_SOURCE_DIR}/fxdiv" CACHE STRING "FXdiv source directory")
 ENDIF()
 
@ -52,7 +47,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading PSimd to ${CONFU_DEPENDENCIES_SOURCE_DIR}/psimd (define PSIMD_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadPSimd.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/psimd-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -97,7 +95,7 @@ IF(NOT DEFINED PSIMD_SOURCE_DIR)
+@@ -100,7 +98,7 @@ IF(NOT DEFINED PSIMD_SOURCE_DIR)
   SET(PSIMD_SOURCE_DIR "${CONFU_DEPENDENCIES_SOURCE_DIR}/psimd" CACHE STRING "PSimd source directory")
 ENDIF()
 
@ -61,7 +56,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading pthreadpool to ${CONFU_DEPENDENCIES_SOURCE_DIR}/pthreadpool (define PTHREADPOOL_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadPThreadPool.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/pthreadpool-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -107,7 +105,7 @@ IF(NOT DEFINED PTHREADPOOL_SOURCE_DIR)
+@@ -110,7 +108,7 @@ IF(NOT DEFINED PTHREADPOOL_SOURCE_DIR)
   SET(PTHREADPOOL_SOURCE_DIR "${CONFU_DEPENDENCIES_SOURCE_DIR}/pthreadpool" CACHE STRING "pthreadpool source directory")
 ENDIF()
 
@ -70,7 +65,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading Google Test to ${CONFU_DEPENDENCIES_SOURCE_DIR}/googletest (define GOOGLETEST_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadGoogleTest.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/googletest-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -117,7 +115,7 @@ IF(QNNPACK_BUILD_TESTS AND NOT DEFINED GOOGLETEST_SOURCE_DIR)
+@@ -120,7 +118,7 @@ IF(QNNPACK_BUILD_TESTS AND NOT DEFINED GOOGLETEST_SOURCE_DIR)
   SET(GOOGLETEST_SOURCE_DIR "${CONFU_DEPENDENCIES_SOURCE_DIR}/googletest" CACHE STRING "Google Test source directory")
 ENDIF()
 
@ -79,7 +74,7 @@ index a5ddc49..05e1279 100644
   MESSAGE(STATUS "Downloading Google Benchmark to ${CONFU_DEPENDENCIES_SOURCE_DIR}/googlebenchmark (define GOOGLEBENCHMARK_SOURCE_DIR to avoid it)")
   CONFIGURE_FILE(cmake/DownloadGoogleBenchmark.cmake "${CONFU_DEPENDENCIES_BINARY_DIR}/googlebenchmark-download/CMakeLists.txt")
   EXECUTE_PROCESS(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
-@@ -229,15 +227,7 @@ IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i[3-6]86|x86_64)$" OR IOS_ARCH MATCHES "^(i
+@@ -232,15 +230,7 @@ IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i[3-6]86|x86_64)$" OR IOS_ARCH MATCHES "^(i
   LIST(APPEND QNNPACK_UKERNELS ${QNNPACK_X86_SSE2_UKERNELS})
 ENDIF()
 
@ -96,89 +91,55 @@ index a5ddc49..05e1279 100644
 SET_TARGET_PROPERTIES(qnnpack PROPERTIES
   C_STANDARD 99
   C_EXTENSIONS YES)
-@@ -272,7 +262,7 @@ TARGET_INCLUDE_DIRECTORIES(qnnpack PRIVATE src)
+@@ -274,6 +264,10 @@ TARGET_INCLUDE_DIRECTORIES(qnnpack PUBLIC include)
+ TARGET_INCLUDE_DIRECTORIES(qnnpack PRIVATE src)
 SET_TARGET_PROPERTIES(qnnpack PROPERTIES PUBLIC_HEADER include/qnnpack.h)
 
- # ---[ Configure clog
-IF(NOT TARGET clog)
-+IF(FALSE)
-   SET(CLOG_BUILD_TESTS OFF CACHE BOOL "")
-   SET(CLOG_RUNTIME_TYPE "${CPUINFO_RUNTIME_TYPE}" CACHE STRING "")
-   ADD_SUBDIRECTORY(
-@@ -281,10 +271,9 @@ IF(NOT TARGET clog)
-   # We build static version of clog but a dynamic library may indirectly depend on it
-   SET_PROPERTY(TARGET clog PROPERTY POSITION_INDEPENDENT_CODE ON)
- ENDIF()
-TARGET_LINK_LIBRARIES(qnnpack PRIVATE clog)
- 
- # ---[ Configure cpuinfo
-IF(NOT TARGET cpuinfo)
-+IF(FALSE)
-   SET(CPUINFO_BUILD_TOOLS OFF CACHE BOOL "")
-   SET(CPUINFO_BUILD_UNIT_TESTS OFF CACHE BOOL "")
-   SET(CPUINFO_BUILD_MOCK_TESTS OFF CACHE BOOL "")
-@@ -293,17 +282,19 @@ IF(NOT TARGET cpuinfo)
-     "${CPUINFO_SOURCE_DIR}"
-     "${CONFU_DEPENDENCIES_BINARY_DIR}/cpuinfo")
- ENDIF()
-TARGET_LINK_LIBRARIES(qnnpack PRIVATE cpuinfo)
-+
 +find_package(cpuinfo CONFIG REQUIRED)
-+target_link_libraries(qnnpack PUBLIC cpuinfo::clog cpuinfo::cpuinfo)
+add_library(cpuinfo ALIAS cpuinfo::cpuinfo)
+add_library(clog ALIAS cpuinfo::clog)
+
+ # ---[ Configure clog
+ IF(NOT TARGET clog)
+   SET(CLOG_BUILD_TESTS OFF CACHE BOOL "")
+@@ -299,6 +293,8 @@ ENDIF()
+ TARGET_LINK_LIBRARIES(qnnpack PRIVATE cpuinfo)
 
 # ---[ Configure pthreadpool
-IF(NOT TARGET pthreadpool)
-+IF(FALSE)
+find_package(unofficial-pthreadpool CONFIG REQUIRED)
+add_library(pthreadpool ALIAS unofficial::pthreadpool)
+ IF(NOT TARGET pthreadpool)
   SET(PTHREADPOOL_BUILD_TESTS OFF CACHE BOOL "")
   SET(PTHREADPOOL_BUILD_BENCHMARKS OFF CACHE BOOL "")
-   ADD_SUBDIRECTORY(
-     "${PTHREADPOOL_SOURCE_DIR}"
-     "${CONFU_DEPENDENCIES_BINARY_DIR}/pthreadpool")
- ENDIF()
-IF(QNNPACK_CUSTOM_THREADPOOL)
-+IF(FALSE)
-   # Depend on pthreadpool interface, but not on implementation.
-   # This is used when QNNPACK user (e.g. Caffe2) provides its own threadpool implementation.
-   TARGET_LINK_LIBRARIES(qnnpack PUBLIC pthreadpool_interface)
-@@ -312,34 +303,38 @@ ELSE()
+@@ -315,6 +311,8 @@ ELSE()
 ENDIF()
 
 # ---[ Configure FXdiv
-IF(NOT TARGET fxdiv)
-+IF(FALSE)
+find_package(unofficial-fxdiv CONFIG REQUIRED)
+add_library(fxdiv ALIAS unofficial::fxdiv::fxdiv)
+ IF(NOT TARGET fxdiv)
   SET(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
   SET(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
-   ADD_SUBDIRECTORY(
-     "${FXDIV_SOURCE_DIR}"
-     "${CONFU_DEPENDENCIES_BINARY_DIR}/fxdiv")
- ENDIF()
-TARGET_LINK_LIBRARIES(qnnpack PRIVATE fxdiv)
-+find_path(FXDIV_INCLUDE_DIRS "fxdiv.h")
-+target_include_directories(qnnpack PRIVATE ${FXDIV_INCLUDE_DIRS})
+@@ -325,6 +323,8 @@ ENDIF()
+ TARGET_LINK_LIBRARIES(qnnpack PRIVATE fxdiv)
 
 # ---[ Configure psimd
-IF(NOT TARGET psimd)
-+IF(FALSE)
+find_package(unofficial-psimd CONFIG REQUIRED)
+add_library(psimd ALIAS unofficial::psimd::psimd)
+ IF(NOT TARGET psimd)
   ADD_SUBDIRECTORY(
     "${PSIMD_SOURCE_DIR}"
-     "${CONFU_DEPENDENCIES_BINARY_DIR}/psimd")
- ENDIF()
-TARGET_LINK_LIBRARIES(qnnpack PRIVATE psimd)
-+find_path(PSIMD_INCLUDE_DIRS "psimd.h")
-+target_include_directories(qnnpack PRIVATE ${PSIMD_INCLUDE_DIRS})
+@@ -333,6 +333,8 @@ ENDIF()
+ TARGET_LINK_LIBRARIES(qnnpack PRIVATE psimd)
 
 # ---[ Configure FP16
-IF(NOT TARGET fp16)
-+IF(FALSE)
+find_package(unofficial-fp16 CONFIG REQUIRED)
+add_library(fp16 ALIAS unofficial::fp16::fp16)
+ IF(NOT TARGET fp16)
   SET(FP16_BUILD_TESTS OFF CACHE BOOL "")
   SET(FP16_BUILD_BENCHMARKS OFF CACHE BOOL "")
-   ADD_SUBDIRECTORY(
-     "${FP16_SOURCE_DIR}"
-     "${CONFU_DEPENDENCIES_BINARY_DIR}/fp16")
- ENDIF()
-TARGET_LINK_LIBRARIES(qnnpack PRIVATE fp16)
-+find_path(FP16_INCLUDE_DIRS "fp16.h")
-+target_include_directories(qnnpack PRIVATE ${FP16_INCLUDE_DIRS})
+@@ -343,6 +345,7 @@ ENDIF()
+ TARGET_LINK_LIBRARIES(qnnpack PRIVATE fp16)
 
 INSTALL(TARGETS qnnpack
 +    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
--- a/ports/qnnpack/vcpkg.json
+++ b/ports/qnnpack/vcpkg.json
@ -1,13 +1,18 @@
 {
  "name": "qnnpack",
  "version-date": "2021-02-26",
-  "port-version": 3,
+  "port-version": 5,
  "description": "Quantized Neural Network PACKage - mobile-optimized implementation of quantized neural network operators",
  "homepage": "https://github.com/pytorch/QNNPACK",
  "license": "BSD-3-Clause",
  "supports": "!windows",
  "dependencies": [
-    "cpuinfo",
+    {
+      "name": "cpuinfo",
+      "features": [
+        "clog"
+      ]
+    },
    "fp16",
    "fxdiv",
    "psimd",
--- a/ports/xnnpack/fix-build.patch
+++ b/ports/xnnpack/fix-build.patch
@ -1,94 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 95df99eae0..7e969cfbb8 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -704,7 +704,7 @@ IF(MSVC)
-     TARGET_COMPILE_DEFINITIONS(operator-utils PRIVATE "restrict=")
-     TARGET_COMPILE_DEFINITIONS(XNNPACK PRIVATE "restrict=")
-   ENDIF()
-
-+  if(0)
-   TARGET_COMPILE_OPTIONS(microkernels-all PRIVATE "$<$<NOT:$<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/fix-cmake.patch
+++ b/ports/xnnpack/fix-cmake.patch
@ -0,0 +1,94 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 856a47d..06f5ea9 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -1040,14 +1040,14 @@ IF(XNNPACK_BUILD_ALL_MICROKERNELS)
+ ENDIF()
+ TARGET_INCLUDE_DIRECTORIES(datatype PRIVATE include src)
+ TARGET_INCLUDE_DIRECTORIES(microkernels-prod PRIVATE include src)
+-TARGET_INCLUDE_DIRECTORIES(hardware-config PRIVATE include src ${CPUINFO_SOURCE_DIR}/include)
+TARGET_INCLUDE_DIRECTORIES(hardware-config PRIVATE include src)
+ TARGET_INCLUDE_DIRECTORIES(indirection PRIVATE include src)
+ TARGET_INCLUDE_DIRECTORIES(microparams-init PRIVATE include src)
+ TARGET_INCLUDE_DIRECTORIES(normalization PRIVATE include src)
+ TARGET_INCLUDE_DIRECTORIES(packing PRIVATE include src)
+ TARGET_INCLUDE_DIRECTORIES(logging PRIVATE include src)
+ IF(XNNPACK_BUILD_LIBRARY)
+-  TARGET_INCLUDE_DIRECTORIES(XNNPACK PUBLIC include)
+  TARGET_INCLUDE_DIRECTORIES(XNNPACK PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include> $<INSTALL_INTERFACE:include>) 
+   TARGET_INCLUDE_DIRECTORIES(XNNPACK PRIVATE src)
+   TARGET_INCLUDE_DIRECTORIES(allocator PRIVATE include src)
+   TARGET_INCLUDE_DIRECTORIES(cache PRIVATE include src)
+@@ -1085,6 +1085,8 @@ IF(XNNPACK_BUILD_WITH_LIBM)
+ ENDIF()
+ 
+ # ---[ Configure cpuinfo
+find_package(cpuinfo CONFIG REQUIRED)
+add_library(cpuinfo ALIAS cpuinfo::cpuinfo)
+ IF(NOT TARGET cpuinfo)
+   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
+     SET(CPUINFO_BUILD_TOOLS OFF CACHE BOOL "")
+@@ -1111,6 +1113,8 @@ IF(XNNPACK_BUILD_LIBRARY)
+ ENDIF()
+ 
+ # ---[ Configure pthreadpool
+find_package(unofficial-pthreadpool CONFIG REQUIRED)
+add_library(pthreadpool ALIAS unofficial::pthreadpool)
+ IF(NOT TARGET pthreadpool)
+   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
+     SET(PTHREADPOOL_BUILD_TESTS OFF CACHE BOOL "")
+@@ -1135,6 +1139,8 @@ ENDIF()
+ TARGET_LINK_LIBRARIES(xnnpack-base INTERFACE pthreadpool)
+ 
+ # ---[ Configure FXdiv
+find_package(unofficial-fxdiv CONFIG REQUIRED)
+add_library(fxdiv ALIAS unofficial::fxdiv::fxdiv)
+ IF(NOT TARGET fxdiv)
+   IF(NOT XNNPACK_USE_SYSTEM_LIBS)
+     SET(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+@@ -1162,14 +1168,27 @@ IF(XNNPACK_BUILD_LIBRARY)
+ ENDIF()
+ 
+ IF(XNNPACK_BUILD_LIBRARY)
+-  INSTALL(TARGETS XNNPACK microkernels-prod
+  INSTALL(TARGETS XNNPACK microkernels-prod xnnpack-base allocator cache hardware-config indirection memory microkernel-utils microparams-init mutex normalization  operators operator-run operator-utils packing subgraph datatype reference-ukernels logging      
+  EXPORT unofficial-xnnpack-config-targets
+      RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+       LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+       ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+       PUBLIC_HEADER DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
+  INSTALL(EXPORT unofficial-xnnpack-config-targets NAMESPACE unofficial::xnnpack::
+    FILE unofficial-xnnpack-config-targets.cmake
+    DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/unofficial-xnnpack) # share/xnnpack
+
+include(CMakePackageConfigHelpers)
+configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in "${CMAKE_CURRENT_BINARY_DIR}/unofficial-xnnpack-config.cmake" INSTALL_DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/unofficial-xnnpack)
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/unofficial-xnnpack-config.cmake DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/unofficial-xnnpack)
+
+
+ ENDIF()
+ 
+ # ---[ Configure KleidiAI
+ IF(XNNPACK_ENABLE_KLEIDIAI)
+  find_package(KleidiAI CONFIG REQUIRED)
+  add_library(kleidiai ALIAS KleidiAI::kleidiai)
+   IF(NOT TARGET kleidiai)
+     IF(NOT XNNPACK_USE_SYSTEM_LIBS)
+       SET(KLEIDIAI_BUILD_TESTS OFF CACHE BOOL "")
+diff --git a/Config.cmake.in b/Config.cmake.in
+new file mode 100644
+index 0000000..8dc60b4
+--- /dev/null
+++ b/Config.cmake.in
+@@ -0,0 +1,11 @@
+@PACKAGE_INIT@
+
+include(CMakeFindDependencyMacro)
+find_dependency(unofficial-fxdiv)
+find_dependency(cpuinfo)
+find_dependency(unofficial-pthreadpool)
+if(@XNNPACK_ENABLE_KLEIDIAI@)
+    find_dependency(KleidiAI)
+endif()
+
+include ( "${CMAKE_CURRENT_LIST_DIR}/unofficial-xnnpack-config-targets.cmake" )
--- a/ports/xnnpack/portfile.cmake
+++ b/ports/xnnpack/portfile.cmake
@ -5,16 +5,18 @@ endif()
 vcpkg_from_github(
    OUT_SOURCE_PATH SOURCE_PATH
    REPO google/XNNPACK
-    REF 51a987591a6fc9f0fc0707077f53d763ac132cbf # 2022-12-22
-    SHA512 1742ad4de3aa1a67cd5a3cbf15d82afc2cc8569a2abac92542f748a440ef37fef4e9d9c302b7b2813e3176934b637c761c1c0956d20ab1ba28b280d9b121d91b
+    REF 51a0103656eff6fc9bfd39a4597923c4b542c883 # 2024-12-03
+    SHA512 ecb859935996416ce51b2b9ffd5a8c3cc5e933dfab96b7b68c8a52e05227ce3416ffa81e8a3fcef8349aa85147aa1c6a07138c7fc8e87a02ad577e1da9126827
    HEAD_REF master
    PATCHES
-        fix-build.patch
+        fix-cmake.patch
 )

+vcpkg_find_acquire_program(PYTHON3)
 vcpkg_cmake_configure(
    SOURCE_PATH "${SOURCE_PATH}"
    OPTIONS
+        "-DPython_EXECUTABLE=${PYTHON3}"
        -DXNNPACK_USE_SYSTEM_LIBS=ON
        -DXNNPACK_ENABLE_ASSEMBLY=ON
        -DXNNPACK_ENABLE_MEMOPT=ON
@ -24,9 +26,7 @@ vcpkg_cmake_configure(
 )
 vcpkg_cmake_install()
 vcpkg_copy_pdbs()
+vcpkg_cmake_config_fixup(PACKAGE_NAME unofficial-${PORT})

 file(INSTALL "${SOURCE_PATH}/LICENSE" DESTINATION "${CURRENT_PACKAGES_DIR}/share/${PORT}" RENAME copyright)
-file(REMOVE_RECURSE "${CURRENT_PACKAGES_DIR}/debug/include"
-                    "${CURRENT_PACKAGES_DIR}/debug/bin"
-                    "${CURRENT_PACKAGES_DIR}/debug/share"
-)
+file(REMOVE_RECURSE "${CURRENT_PACKAGES_DIR}/debug/include")
--- a/ports/xnnpack/vcpkg.json
+++ b/ports/xnnpack/vcpkg.json
@ -1,18 +1,31 @@
 {
  "name": "xnnpack",
-  "version-date": "2022-12-22",
+  "version-date": "2024-08-20",
  "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 & !arm32",
  "dependencies": [
-    "cpuinfo",
+    {
+      "name": "cpuinfo",
+      "features": [
+        "clog"
+      ]
+    },
    "fp16",
    "fxdiv",
+    {
+      "name": "kleidiai",
+      "platform": "arm64"
+    },
    "pthreadpool",
    {
      "name": "vcpkg-cmake",
      "host": true
+    },
+    {
+      "name": "vcpkg-cmake-config",
+      "host": true
    }
  ]
 }
--- a/scripts/ci.feature.baseline.txt
+++ b/scripts/ci.feature.baseline.txt
@ -1333,8 +1333,7 @@ librsvg:x64-uwp = cascade
 libsm:arm64-uwp = cascade
 libsm:x64-uwp = cascade
 libtorch:x86-windows = cascade
-libtorch[vulkan] = cascade # no vulkan sdk installed
-libtorch(uwp | (arm & windows) | (arm64 & osx)) = cascade
+libtorch(uwp | (arm & windows)) = cascade
 libwebp[all,vwebp](uwp) = cascade
 libwebp[vwebp](android) = cascade
 libwebp[vwebp-sdl](android | osx | uwp | (windows & static)) = cascade
--- a/versions/baseline.json
+++ b/versions/baseline.json
@ -2133,8 +2133,8 @@
      "port-version": 0
    },
    "cpuinfo": {
-      "baseline": "2022-07-19",
-      "port-version": 3
+      "baseline": "2025-07-25",
+      "port-version": 0
    },
    "cr": {
      "baseline": "2020-04-26",
@ -2866,7 +2866,7 @@
    },
    "fbgemm": {
      "baseline": "1.0.0",
-      "port-version": 0
+      "port-version": 1
    },
    "fbthrift": {
      "baseline": "2025.05.19.00",
@ -5509,8 +5509,8 @@
      "port-version": 2
    },
    "libtorch": {
-      "baseline": "2.1.2",
-      "port-version": 12
+      "baseline": "2.7.1",
+      "port-version": 0
    },
    "libtorrent": {
      "baseline": "2.0.11",
@ -7754,7 +7754,7 @@
    },
    "qnnpack": {
      "baseline": "2021-02-26",
-      "port-version": 3
+      "port-version": 5
    },
    "qoi": {
      "baseline": "2023-08-10",
@ -10521,7 +10521,7 @@
      "port-version": 1
    },
    "xnnpack": {
-      "baseline": "2022-12-22",
+      "baseline": "2024-08-20",
      "port-version": 0
    },
    "xorg-macros": {
--- a/versions/c-/cpuinfo.json
+++ b/versions/c-/cpuinfo.json
@ -1,5 +1,10 @@
 {
  "versions": [
+    {
+      "git-tree": "1a3c239a30cc457f10b8d6cca995bc2c7db9b150",
+      "version-date": "2025-07-25",
+      "port-version": 0
+    },
    {
      "git-tree": "f1b7e10e1de1dcd09118adc19c5bc831417dcc8c",
      "version-date": "2022-07-19",
--- a/versions/f-/fbgemm.json
+++ b/versions/f-/fbgemm.json
@ -1,5 +1,10 @@
 {
  "versions": [
+    {
+      "git-tree": "86dc5f7bb07a7e785aef6627e4af57f51c9c87d0",
+      "version": "1.0.0",
+      "port-version": 1
+    },
    {
      "git-tree": "e055c95b9158e871ada9e49d4dfc92c618f995f2",
      "version": "1.0.0",
--- a/versions/l-/libtorch.json
+++ b/versions/l-/libtorch.json
@ -1,5 +1,10 @@
 {
  "versions": [
+    {
+      "git-tree": "b79f7f4e0fc9c9dbfe7528a0c7c96ee8f811f753",
+      "version": "2.7.1",
+      "port-version": 0
+    },
    {
      "git-tree": "42cff2032d0a7670edde067c822897cc6be1cf85",
      "version": "2.1.2",
--- a/versions/q-/qnnpack.json
+++ b/versions/q-/qnnpack.json
@ -1,5 +1,15 @@
 {
  "versions": [
+    {
+      "git-tree": "3066be09481d88681b8a20480053585f173c9821",
+      "version-date": "2021-02-26",
+      "port-version": 5
+    },
+    {
+      "git-tree": "6cfef60767e1bc650df581901e4d14e132d5fdea",
+      "version-date": "2021-02-26",
+      "port-version": 4
+    },
    {
      "git-tree": "cee68bf91e46cc783fe3f03d4525e02c0793389c",
      "version-date": "2021-02-26",
--- a/versions/x-/xnnpack.json
+++ b/versions/x-/xnnpack.json
@ -1,5 +1,10 @@
 {
  "versions": [
+    {
+      "git-tree": "8f49750bf854bc25fa46d069eca34aece7a5867f",
+      "version-date": "2024-08-20",
+      "port-version": 0
+    },
    {
      "git-tree": "7a84132ceac551e5ba3da8117fdc23f4f1571fd2",
      "version-date": "2022-12-22",