Update (base update)

[ghstack-poisoned]
2025-11-11 22:34:53 +08:00 · 2024-10-20 02:09:32 +08:00
parent 9ca1b1ed90 2f6a70bfea
commit f656f68880
487 changed files with 7261 additions and 3811 deletions
--- a/.ci/docker/requirements-ci.txt
+++ b/.ci/docker/requirements-ci.txt
@ -363,3 +363,8 @@ setuptools

 ninja==1.11.1 ; platform_machine == "aarch64"
 scons==4.5.2 ; platform_machine == "aarch64"
+
+pulp==2.9.0 ; python_version >= "3.8"
+#Description: required for testing ilp formulaiton under torch/distributed/_tools
+#Pinned versions: 2.9.0
+#test that import: test_sac_ilp.py
--- a/.ci/libtorch/build.sh
+++ b/.ci/libtorch/build.sh
@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+# This is mostly just a shim to manywheel/build.sh
+# TODO: Make this a dedicated script to build just libtorch
+
+set -ex
+
+SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
+
+USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh
--- a/.ci/manywheel/LICENSE
+++ b/.ci/manywheel/LICENSE
@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2016 manylinux
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
--- a/.ci/manywheel/build.sh
+++ b/.ci/manywheel/build.sh
@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+
+set -ex
+
+SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
+
+case "${GPU_ARCH_TYPE:-BLANK}" in
+    BLANK)
+        # Legacy behavior for CircleCI
+        bash "${SCRIPTPATH}/build_cuda.sh"
+        ;;
+    cuda)
+        bash "${SCRIPTPATH}/build_cuda.sh"
+        ;;
+    rocm)
+        bash "${SCRIPTPATH}/build_rocm.sh"
+        ;;
+    cpu | cpu-cxx11-abi | cpu-s390x | xpu)
+        bash "${SCRIPTPATH}/build_cpu.sh"
+        ;;
+    *)
+        echo "Un-recognized GPU_ARCH_TYPE '${GPU_ARCH_TYPE}', exiting..."
+        exit 1
+        ;;
+esac
--- a/.ci/manywheel/build_common.sh
+++ b/.ci/manywheel/build_common.sh
@ -0,0 +1,505 @@
+#!/usr/bin/env bash
+# meant to be called only from the neighboring build.sh and build_cpu.sh scripts
+
+set -ex
+SOURCE_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null && pwd )"
+
+
+# Require only one python installation
+if [[ -z "$DESIRED_PYTHON" ]]; then
+    echo "Need to set DESIRED_PYTHON env variable"
+    exit 1
+fi
+if [[ -n "$BUILD_PYTHONLESS" && -z "$LIBTORCH_VARIANT" ]]; then
+    echo "BUILD_PYTHONLESS is set, so need LIBTORCH_VARIANT to also be set"
+    echo "LIBTORCH_VARIANT should be one of shared-with-deps shared-without-deps static-with-deps static-without-deps"
+    exit 1
+fi
+
+# Function to retry functions that sometimes timeout or have flaky failures
+retry () {
+    $*  || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
+}
+
+# TODO move this into the Docker images
+OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
+    retry dnf install -q -y zip openssl
+elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
+    # TODO: Remove this once nvidia package repos are back online
+    # Comment out nvidia repositories to prevent them from getting apt-get updated, see https://github.com/pytorch/pytorch/issues/74968
+    # shellcheck disable=SC2046
+    sed -i 's/.*nvidia.*/# &/' $(find /etc/apt/ -type f -name "*.list")
+
+    retry apt-get update
+    retry apt-get -y install zip openssl
+fi
+
+# We use the package name to test the package by passing this to 'pip install'
+# This is the env variable that setup.py uses to name the package. Note that
+# pip 'normalizes' the name first by changing all - to _
+if [[ -z "$TORCH_PACKAGE_NAME" ]]; then
+    TORCH_PACKAGE_NAME='torch'
+fi
+
+if [[ -z "$TORCH_NO_PYTHON_PACKAGE_NAME" ]]; then
+    TORCH_NO_PYTHON_PACKAGE_NAME='torch_no_python'
+fi
+
+TORCH_PACKAGE_NAME="$(echo $TORCH_PACKAGE_NAME | tr '-' '_')"
+TORCH_NO_PYTHON_PACKAGE_NAME="$(echo $TORCH_NO_PYTHON_PACKAGE_NAME | tr '-' '_')"
+echo "Expecting the built wheels to all be called '$TORCH_PACKAGE_NAME' or '$TORCH_NO_PYTHON_PACKAGE_NAME'"
+
+# Version: setup.py uses $PYTORCH_BUILD_VERSION.post$PYTORCH_BUILD_NUMBER if
+# PYTORCH_BUILD_NUMBER > 1
+build_version="$PYTORCH_BUILD_VERSION"
+build_number="$PYTORCH_BUILD_NUMBER"
+if [[ -n "$OVERRIDE_PACKAGE_VERSION" ]]; then
+    # This will be the *exact* version, since build_number<1
+    build_version="$OVERRIDE_PACKAGE_VERSION"
+    build_number=0
+fi
+if [[ -z "$build_version" ]]; then
+    build_version=1.0.0
+fi
+if [[ -z "$build_number" ]]; then
+    build_number=1
+fi
+export PYTORCH_BUILD_VERSION=$build_version
+export PYTORCH_BUILD_NUMBER=$build_number
+
+export CMAKE_LIBRARY_PATH="/opt/intel/lib:/lib:$CMAKE_LIBRARY_PATH"
+export CMAKE_INCLUDE_PATH="/opt/intel/include:$CMAKE_INCLUDE_PATH"
+
+if [[ -e /opt/openssl ]]; then
+    export OPENSSL_ROOT_DIR=/opt/openssl
+    export CMAKE_INCLUDE_PATH="/opt/openssl/include":$CMAKE_INCLUDE_PATH
+fi
+
+# If given a python version like 3.6m or 2.7mu, convert this to the format we
+# expect. The binary CI jobs pass in python versions like this; they also only
+# ever pass one python version, so we assume that DESIRED_PYTHON is not a list
+# in this case
+if [[ -n "$DESIRED_PYTHON" && $DESIRED_PYTHON =~ ([0-9].[0-9]+)t ]]; then
+    python_digits="$(echo $DESIRED_PYTHON | tr -cd [:digit:])"
+    py_majmin="${DESIRED_PYTHON}"
+    DESIRED_PYTHON="cp${python_digits}-cp${python_digits}t"
+elif [[ -n "$DESIRED_PYTHON" && "$DESIRED_PYTHON" != cp* ]]; then
+    python_nodot="$(echo $DESIRED_PYTHON | tr -d m.u)"
+    DESIRED_PYTHON="cp${python_nodot}-cp${python_nodot}"
+    if [[ ${python_nodot} -ge 310 ]]; then
+        py_majmin="${DESIRED_PYTHON:2:1}.${DESIRED_PYTHON:3:2}"
+    else
+        py_majmin="${DESIRED_PYTHON:2:1}.${DESIRED_PYTHON:3:1}"
+    fi
+fi
+
+pydir="/opt/python/$DESIRED_PYTHON"
+export PATH="$pydir/bin:$PATH"
+echo "Will build for Python version: ${DESIRED_PYTHON} with ${python_installation}"
+
+mkdir -p /tmp/$WHEELHOUSE_DIR
+
+export PATCHELF_BIN=/usr/local/bin/patchelf
+patchelf_version=$($PATCHELF_BIN --version)
+echo "patchelf version: " $patchelf_version
+if [[ "$patchelf_version" == "patchelf 0.9" ]]; then
+    echo "Your patchelf version is too old. Please use version >= 0.10."
+    exit 1
+fi
+
+########################################################
+# Compile wheels as well as libtorch
+#######################################################
+if [[ -z "$PYTORCH_ROOT" ]]; then
+    echo "Need to set PYTORCH_ROOT env variable"
+    exit 1
+fi
+pushd "$PYTORCH_ROOT"
+python setup.py clean
+retry pip install -qr requirements.txt
+case ${DESIRED_PYTHON} in
+  cp31*)
+    retry pip install -q --pre numpy==2.1.0
+    ;;
+  # Should catch 3.9+
+  *)
+    retry pip install -q --pre numpy==2.0.2
+    ;;
+esac
+
+if [[ "$DESIRED_DEVTOOLSET" == *"cxx11-abi"* ]]; then
+    export _GLIBCXX_USE_CXX11_ABI=1
+else
+    export _GLIBCXX_USE_CXX11_ABI=0
+fi
+
+if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+    echo "Calling build_amd.py at $(date)"
+    python tools/amd_build/build_amd.py
+fi
+
+# This value comes from binary_linux_build.sh (and should only be set to true
+# for master / release branches)
+BUILD_DEBUG_INFO=${BUILD_DEBUG_INFO:=0}
+
+if [[ $BUILD_DEBUG_INFO == "1" ]]; then
+    echo "Building wheel and debug info"
+else
+    echo "BUILD_DEBUG_INFO was not set, skipping debug info"
+fi
+
+if [[ "$DISABLE_RCCL" = 1 ]]; then
+    echo "Disabling NCCL/RCCL in pyTorch"
+    USE_RCCL=0
+    USE_NCCL=0
+    USE_KINETO=0
+else
+    USE_RCCL=1
+    USE_NCCL=1
+    USE_KINETO=1
+fi
+
+echo "Calling setup.py bdist at $(date)"
+
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 \
+    BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+    USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+    python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 \
+    BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+    USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+    python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR --cmake
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
+        EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+        BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+        USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+        python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+fi
+echo "Finished setup.py bdist at $(date)"
+
+# Build libtorch packages
+if [[ -n "$BUILD_PYTHONLESS" ]]; then
+    # Now build pythonless libtorch
+    # Note - just use whichever python we happen to be on
+    python setup.py clean
+
+    if [[ $LIBTORCH_VARIANT = *"static"* ]]; then
+        STATIC_CMAKE_FLAG="-DTORCH_STATIC=1"
+    fi
+
+    mkdir -p build
+    pushd build
+    echo "Calling tools/build_libtorch.py at $(date)"
+    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
+         EXTRA_CAFFE2_CMAKE_FLAGS="${EXTRA_CAFFE2_CMAKE_FLAGS[@]} $STATIC_CMAKE_FLAG" \
+         python ../tools/build_libtorch.py
+    echo "Finished tools/build_libtorch.py at $(date)"
+    popd
+
+    mkdir -p libtorch/{lib,bin,include,share}
+    cp -r build/build/lib libtorch/
+
+    # for now, the headers for the libtorch package will just be copied in
+    # from one of the wheels (this is from when this script built multiple
+    # wheels at once)
+    ANY_WHEEL=$(ls /tmp/$WHEELHOUSE_DIR/torch*.whl | head -n1)
+    unzip -d any_wheel $ANY_WHEEL
+    if [[ -d any_wheel/torch/include ]]; then
+        cp -r any_wheel/torch/include libtorch/
+    else
+        cp -r any_wheel/torch/lib/include libtorch/
+    fi
+    cp -r any_wheel/torch/share/cmake libtorch/share/
+    rm -rf any_wheel
+
+    echo $PYTORCH_BUILD_VERSION > libtorch/build-version
+    echo "$(pushd $PYTORCH_ROOT && git rev-parse HEAD)" > libtorch/build-hash
+
+    mkdir -p /tmp/$LIBTORCH_HOUSE_DIR
+
+    if [[ "$DESIRED_DEVTOOLSET" == *"cxx11-abi"* ]]; then
+        LIBTORCH_ABI="cxx11-abi-"
+    else
+        LIBTORCH_ABI=
+    fi
+
+    zip -rq /tmp/$LIBTORCH_HOUSE_DIR/libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-$PYTORCH_BUILD_VERSION.zip libtorch
+    cp /tmp/$LIBTORCH_HOUSE_DIR/libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-$PYTORCH_BUILD_VERSION.zip \
+       /tmp/$LIBTORCH_HOUSE_DIR/libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-latest.zip
+fi
+
+popd
+
+#######################################################################
+# ADD DEPENDENCIES INTO THE WHEEL
+#
+# auditwheel repair doesn't work correctly and is buggy
+# so manually do the work of copying dependency libs and patchelfing
+# and fixing RECORDS entries correctly
+######################################################################
+
+fname_with_sha256() {
+    HASH=$(sha256sum $1 | cut -c1-8)
+    DIRNAME=$(dirname $1)
+    BASENAME=$(basename $1)
+    # Do not rename nvrtc-builtins.so as they are dynamically loaded
+    # by libnvrtc.so
+    # Similarly don't mangle libcudnn and libcublas library names
+    if [[ $BASENAME == "libnvrtc-builtins.s"* || $BASENAME == "libcudnn"* || $BASENAME == "libcublas"*  ]]; then
+        echo $1
+    else
+        INITNAME=$(echo $BASENAME | cut -f1 -d".")
+        ENDNAME=$(echo $BASENAME | cut -f 2- -d".")
+        echo "$DIRNAME/$INITNAME-$HASH.$ENDNAME"
+    fi
+}
+
+fname_without_so_number() {
+    LINKNAME=$(echo $1 | sed -e 's/\.so.*/.so/g')
+    echo "$LINKNAME"
+}
+
+make_wheel_record() {
+    FPATH=$1
+    if echo $FPATH | grep RECORD >/dev/null 2>&1; then
+        # if the RECORD file, then
+        echo "$FPATH,,"
+    else
+        HASH=$(openssl dgst -sha256 -binary $FPATH | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
+        FSIZE=$(ls -nl $FPATH | awk '{print $5}')
+        echo "$FPATH,sha256=$HASH,$FSIZE"
+    fi
+}
+
+replace_needed_sofiles() {
+    find $1 -name '*.so*' | while read sofile; do
+        origname=$2
+        patchedname=$3
+        if [[ "$origname" != "$patchedname" ]] || [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+            set +e
+            origname=$($PATCHELF_BIN --print-needed $sofile | grep "$origname.*")
+            ERRCODE=$?
+            set -e
+            if [ "$ERRCODE" -eq "0" ]; then
+                echo "patching $sofile entry $origname to $patchedname"
+                $PATCHELF_BIN --replace-needed $origname $patchedname $sofile
+            fi
+        fi
+    done
+}
+
+echo 'Built this wheel:'
+ls /tmp/$WHEELHOUSE_DIR
+mkdir -p "/$WHEELHOUSE_DIR"
+mv /tmp/$WHEELHOUSE_DIR/torch*linux*.whl /$WHEELHOUSE_DIR/
+
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    mv /tmp/$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/ || true
+fi
+
+if [[ -n "$BUILD_PYTHONLESS" ]]; then
+    mkdir -p /$LIBTORCH_HOUSE_DIR
+    mv /tmp/$LIBTORCH_HOUSE_DIR/*.zip /$LIBTORCH_HOUSE_DIR
+    rm -rf /tmp/$LIBTORCH_HOUSE_DIR
+fi
+rm -rf /tmp/$WHEELHOUSE_DIR
+rm -rf /tmp_dir
+mkdir /tmp_dir
+pushd /tmp_dir
+
+for pkg in /$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/torch*linux*.whl /$LIBTORCH_HOUSE_DIR/libtorch*.zip; do
+
+    # if the glob didn't match anything
+    if [[ ! -e $pkg ]]; then
+        continue
+    fi
+
+    rm -rf tmp
+    mkdir -p tmp
+    cd tmp
+    cp $pkg .
+
+    unzip -q $(basename $pkg)
+    rm -f $(basename $pkg)
+
+    if [[ -d torch ]]; then
+        PREFIX=torch
+    else
+        PREFIX=libtorch
+    fi
+
+    if [[ $pkg != *"without-deps"* ]]; then
+        # copy over needed dependent .so files over and tag them with their hash
+        patched=()
+        for filepath in "${DEPS_LIST[@]}"; do
+            filename=$(basename $filepath)
+            destpath=$PREFIX/lib/$filename
+            if [[ "$filepath" != "$destpath" ]]; then
+                cp $filepath $destpath
+            fi
+
+            # ROCm workaround for roctracer dlopens
+            if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+                patchedpath=$(fname_without_so_number $destpath)
+            # Keep the so number for XPU dependencies
+            elif [[ "$DESIRED_CUDA" == *"xpu"* ]]; then
+                patchedpath=$destpath
+            else
+                patchedpath=$(fname_with_sha256 $destpath)
+            fi
+            patchedname=$(basename $patchedpath)
+            if [[ "$destpath" != "$patchedpath" ]]; then
+                mv $destpath $patchedpath
+            fi
+            patched+=("$patchedname")
+            echo "Copied $filepath to $patchedpath"
+        done
+
+        echo "patching to fix the so names to the hashed names"
+        for ((i=0;i<${#DEPS_LIST[@]};++i)); do
+            replace_needed_sofiles $PREFIX ${DEPS_SONAME[i]} ${patched[i]}
+            # do the same for caffe2, if it exists
+            if [[ -d caffe2 ]]; then
+                replace_needed_sofiles caffe2 ${DEPS_SONAME[i]} ${patched[i]}
+            fi
+        done
+
+        # copy over needed auxiliary files
+        for ((i=0;i<${#DEPS_AUX_SRCLIST[@]};++i)); do
+            srcpath=${DEPS_AUX_SRCLIST[i]}
+            dstpath=$PREFIX/${DEPS_AUX_DSTLIST[i]}
+            mkdir -p $(dirname $dstpath)
+            cp $srcpath $dstpath
+        done
+    fi
+
+    # set RPATH of _C.so and similar to $ORIGIN, $ORIGIN/lib
+    find $PREFIX -maxdepth 1 -type f -name "*.so*" | while read sofile; do
+        echo "Setting rpath of $sofile to ${C_SO_RPATH:-'$ORIGIN:$ORIGIN/lib'}"
+        $PATCHELF_BIN --set-rpath ${C_SO_RPATH:-'$ORIGIN:$ORIGIN/lib'} ${FORCE_RPATH:-} $sofile
+        $PATCHELF_BIN --print-rpath $sofile
+    done
+
+    # set RPATH of lib/ files to $ORIGIN
+    find $PREFIX/lib -maxdepth 1 -type f -name "*.so*" | while read sofile; do
+        echo "Setting rpath of $sofile to ${LIB_SO_RPATH:-'$ORIGIN'}"
+        $PATCHELF_BIN --set-rpath ${LIB_SO_RPATH:-'$ORIGIN'} ${FORCE_RPATH:-} $sofile
+        $PATCHELF_BIN --print-rpath $sofile
+    done
+
+    # regenerate the RECORD file with new hashes
+    record_file=$(echo $(basename $pkg) | sed -e 's/-cp.*$/.dist-info\/RECORD/g')
+    if [[ -e $record_file ]]; then
+        echo "Generating new record file $record_file"
+        : > "$record_file"
+        # generate records for folders in wheel
+        find * -type f | while read fname; do
+            make_wheel_record "$fname" >>"$record_file"
+        done
+    fi
+
+    if [[ $BUILD_DEBUG_INFO == "1" ]]; then
+        pushd "$PREFIX/lib"
+
+        # Duplicate library into debug lib
+        cp libtorch_cpu.so libtorch_cpu.so.dbg
+
+        # Keep debug symbols on debug lib
+        strip --only-keep-debug libtorch_cpu.so.dbg
+
+        # Remove debug info from release lib
+        strip --strip-debug libtorch_cpu.so
+
+        objcopy libtorch_cpu.so --add-gnu-debuglink=libtorch_cpu.so.dbg
+
+        # Zip up debug info
+        mkdir -p /tmp/debug
+        mv libtorch_cpu.so.dbg /tmp/debug/libtorch_cpu.so.dbg
+        CRC32=$(objcopy --dump-section .gnu_debuglink=>(tail -c4 | od -t x4 -An | xargs echo) libtorch_cpu.so)
+
+        pushd /tmp
+        PKG_NAME=$(basename "$pkg" | sed 's/\.whl$//g')
+        zip /tmp/debug-whl-libtorch-"$PKG_NAME"-"$CRC32".zip /tmp/debug/libtorch_cpu.so.dbg
+        cp /tmp/debug-whl-libtorch-"$PKG_NAME"-"$CRC32".zip "$PYTORCH_FINAL_PACKAGE_DIR"
+        popd
+
+        popd
+    fi
+
+    # zip up the wheel back
+    zip -rq $(basename $pkg) $PREIX*
+
+    # replace original wheel
+    rm -f $pkg
+    mv $(basename $pkg) $pkg
+    cd ..
+    rm -rf tmp
+done
+
+# Copy wheels to host machine for persistence before testing
+if [[ -n "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
+    mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
+    if [[ -n "$BUILD_PYTHONLESS" ]]; then
+        cp /$LIBTORCH_HOUSE_DIR/libtorch*.zip "$PYTORCH_FINAL_PACKAGE_DIR"
+    else
+        cp /$WHEELHOUSE_DIR/torch*.whl "$PYTORCH_FINAL_PACKAGE_DIR"
+    fi
+fi
+
+# remove stuff before testing
+rm -rf /opt/rh
+if ls /usr/local/cuda* >/dev/null 2>&1; then
+    rm -rf /usr/local/cuda*
+fi
+
+
+# Test that all the wheels work
+if [[ -z "$BUILD_PYTHONLESS" ]]; then
+  export OMP_NUM_THREADS=4 # on NUMA machines this takes too long
+  pushd $PYTORCH_ROOT/test
+
+  # Install the wheel for this Python version
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip uninstall -y "$TORCH_NO_PYTHON_PACKAGE_NAME" || true
+  fi
+
+  pip uninstall -y "$TORCH_PACKAGE_NAME"
+
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip install "$TORCH_NO_PYTHON_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
+  fi
+
+  pip install "$TORCH_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
+
+  # Print info on the libraries installed in this wheel
+  # Rather than adjust find command to skip non-library files with an embedded *.so* in their name,
+  # since this is only for reporting purposes, we add the || true to the ldd command.
+  installed_libraries=($(find "$pydir/lib/python${py_majmin}/site-packages/torch/" -name '*.so*'))
+  echo "The wheel installed all of the libraries: ${installed_libraries[@]}"
+  for installed_lib in "${installed_libraries[@]}"; do
+      ldd "$installed_lib" || true
+  done
+
+  # Run the tests
+  echo "$(date) :: Running tests"
+  pushd "$PYTORCH_ROOT"
+
+  #TODO: run_tests.sh and check_binary.sh should be moved to pytorch/pytorch project
+  LD_LIBRARY_PATH=/usr/local/nvidia/lib64 \
+          "/builder/run_tests.sh" manywheel "${py_majmin}" "$DESIRED_CUDA"
+  popd
+  echo "$(date) :: Finished tests"
+fi
--- a/.ci/manywheel/build_cpu.sh
+++ b/.ci/manywheel/build_cpu.sh
@ -0,0 +1,99 @@
+#!/usr/bin/env bash
+
+set -ex
+
+GPU_ARCH_TYPE=${GPU_ARCH_TYPE:-cpu}
+
+export TH_BINARY_BUILD=1
+export USE_CUDA=0
+
+# Keep an array of cmake variables to add to
+if [[ -z "$CMAKE_ARGS" ]]; then
+    # These are passed to tools/build_pytorch_libs.sh::build()
+    CMAKE_ARGS=()
+fi
+if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
+    # These are passed to tools/build_pytorch_libs.sh::build_caffe2()
+    EXTRA_CAFFE2_CMAKE_FLAGS=()
+fi
+
+DIR_SUFFIX=cpu
+if [[ "$GPU_ARCH_TYPE" == "xpu" ]]; then
+    DIR_SUFFIX=xpu
+    # Refer https://www.intel.com/content/www/us/en/developer/articles/tool/pytorch-prerequisites-for-intel-gpu/2-5.html
+    source /opt/intel/oneapi/pytorch-gpu-dev-0.5/oneapi-vars.sh
+    source /opt/intel/oneapi/pti/latest/env/vars.sh
+    export USE_STATIC_MKL=1
+fi
+
+WHEELHOUSE_DIR="wheelhouse$DIR_SUFFIX"
+LIBTORCH_HOUSE_DIR="libtorch_house$DIR_SUFFIX"
+if [[ -z "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
+    if [[ -z "$BUILD_PYTHONLESS" ]]; then
+        PYTORCH_FINAL_PACKAGE_DIR="/remote/wheelhouse$DIR_SUFFIX"
+    else
+        PYTORCH_FINAL_PACKAGE_DIR="/remote/libtorch_house$DIR_SUFFIX"
+    fi
+fi
+mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
+
+OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
+    if [[ "$(uname -m)" == "s390x" ]]; then
+        LIBGOMP_PATH="/usr/lib/s390x-linux-gnu/libgomp.so.1"
+    else
+        LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
+    fi
+fi
+
+DEPS_LIST=(
+    "$LIBGOMP_PATH"
+)
+
+DEPS_SONAME=(
+    "libgomp.so.1"
+)
+
+if [[ "$GPU_ARCH_TYPE" == "xpu" ]]; then
+    echo "Bundling with xpu support package libs."
+    DEPS_LIST+=(
+        "/opt/intel/oneapi/compiler/latest/lib/libsycl-preview.so.7"
+        "/opt/intel/oneapi/compiler/latest/lib/libOpenCL.so.1"
+        "/opt/intel/oneapi/compiler/latest/lib/libxptifw.so"
+        "/opt/intel/oneapi/compiler/latest/lib/libsvml.so"
+        "/opt/intel/oneapi/compiler/latest/lib/libirng.so"
+        "/opt/intel/oneapi/compiler/latest/lib/libimf.so"
+        "/opt/intel/oneapi/compiler/latest/lib/libintlc.so.5"
+        "/opt/intel/oneapi/compiler/latest/lib/libpi_level_zero.so"
+        "/opt/intel/oneapi/pti/latest/lib/libpti_view.so.0.9"
+        "/opt/intel/oneapi/pti/latest/lib/libpti.so.0.9"
+    )
+    DEPS_SONAME+=(
+        "libsycl-preview.so.7"
+        "libOpenCL.so.1"
+        "libxptifw.so"
+        "libsvml.so"
+        "libirng.so"
+        "libimf.so"
+        "libintlc.so.5"
+        "libpi_level_zero.so"
+        "libpti_view.so.0.9"
+        "libpti.so.0.9"
+    )
+fi
+
+rm -rf /usr/local/cuda*
+
+SOURCE_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null && pwd )"
+if [[ -z "$BUILD_PYTHONLESS" ]]; then
+    BUILD_SCRIPT=build_common.sh
+else
+    BUILD_SCRIPT=build_libtorch.sh
+fi
+source ${SOURCE_DIR}/${BUILD_SCRIPT}
--- a/.ci/manywheel/build_cuda.sh
+++ b/.ci/manywheel/build_cuda.sh
@ -0,0 +1,290 @@
+#!/usr/bin/env bash
+
+set -ex
+
+SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P ))"
+
+export TORCH_NVCC_FLAGS="-Xfatbin -compress-all"
+export NCCL_ROOT_DIR=/usr/local/cuda
+export TH_BINARY_BUILD=1
+export USE_STATIC_CUDNN=1
+export USE_STATIC_NCCL=1
+export ATEN_STATIC_CUDA=1
+export USE_CUDA_STATIC_LINK=1
+export INSTALL_TEST=0 # dont install test binaries into site-packages
+export USE_CUPTI_SO=0
+export USE_CUSPARSELT=${USE_CUSPARSELT:-1} # Enable if not disabled by libtorch build
+
+# Keep an array of cmake variables to add to
+if [[ -z "$CMAKE_ARGS" ]]; then
+    # These are passed to tools/build_pytorch_libs.sh::build()
+    CMAKE_ARGS=()
+fi
+if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
+    # These are passed to tools/build_pytorch_libs.sh::build_caffe2()
+    EXTRA_CAFFE2_CMAKE_FLAGS=()
+fi
+
+# Determine CUDA version and architectures to build for
+#
+# NOTE: We should first check `DESIRED_CUDA` when determining `CUDA_VERSION`,
+# because in some cases a single Docker image can have multiple CUDA versions
+# on it, and `nvcc --version` might not show the CUDA version we want.
+if [[ -n "$DESIRED_CUDA" ]]; then
+    # If the DESIRED_CUDA already matches the format that we expect
+    if [[ ${DESIRED_CUDA} =~ ^[0-9]+\.[0-9]+$ ]]; then
+        CUDA_VERSION=${DESIRED_CUDA}
+    else
+        # cu90, cu92, cu100, cu101
+        if [[ ${#DESIRED_CUDA} -eq 4 ]]; then
+            CUDA_VERSION="${DESIRED_CUDA:2:1}.${DESIRED_CUDA:3:1}"
+        elif [[ ${#DESIRED_CUDA} -eq 5 ]]; then
+            CUDA_VERSION="${DESIRED_CUDA:2:2}.${DESIRED_CUDA:4:1}"
+        fi
+    fi
+    echo "Using CUDA $CUDA_VERSION as determined by DESIRED_CUDA"
+
+    # There really has to be a better way to do this - eli
+    # Possibly limiting builds to specific cuda versions be delimiting images would be a choice
+    if [[ "$OS_NAME" == *"Ubuntu"* ]]; then
+        echo "Switching to CUDA version ${DESIRED_CUDA}"
+        /builder/conda/switch_cuda_version.sh "${DESIRED_CUDA}"
+    fi
+else
+    CUDA_VERSION=$(nvcc --version|grep release|cut -f5 -d" "|cut -f1 -d",")
+    echo "CUDA $CUDA_VERSION Detected"
+fi
+
+cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
+
+TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6"
+case ${CUDA_VERSION} in
+    12.4)
+        if [[ "$GPU_ARCH_TYPE" = "cuda-aarch64" ]]; then
+            TORCH_CUDA_ARCH_LIST="9.0"
+        else
+            TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0+PTX"
+        fi
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    12.1)
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    11.8)
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};3.7;9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    11.[67])
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};3.7"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    *)
+        echo "unknown cuda version $CUDA_VERSION"
+        exit 1
+        ;;
+esac
+
+export TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
+echo "${TORCH_CUDA_ARCH_LIST}"
+
+# Package directories
+WHEELHOUSE_DIR="wheelhouse$cuda_version_nodot"
+LIBTORCH_HOUSE_DIR="libtorch_house$cuda_version_nodot"
+if [[ -z "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
+    if [[ -z "$BUILD_PYTHONLESS" ]]; then
+        PYTORCH_FINAL_PACKAGE_DIR="/remote/wheelhouse$cuda_version_nodot"
+    else
+        PYTORCH_FINAL_PACKAGE_DIR="/remote/libtorch_house$cuda_version_nodot"
+    fi
+fi
+mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
+
+OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
+    LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
+fi
+
+DEPS_LIST=(
+    "$LIBGOMP_PATH"
+)
+DEPS_SONAME=(
+    "libgomp.so.1"
+)
+
+if [[ $USE_CUSPARSELT == "1" ]]; then
+        DEPS_SONAME+=(
+            "libcusparseLt.so.0"
+        )
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcusparseLt.so.0"
+        )
+fi
+
+if [[ $CUDA_VERSION == "12.1" || $CUDA_VERSION == "12.4" ]]; then
+    export USE_STATIC_CUDNN=0
+    # Try parallelizing nvcc as well
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+
+    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
+        echo "Bundling with cudnn and cublas."
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcudnn_adv.so.9"
+            "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
+            "/usr/local/cuda/lib64/libcudnn_graph.so.9"
+            "/usr/local/cuda/lib64/libcudnn_ops.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
+            "/usr/local/cuda/lib64/libcudnn.so.9"
+            "/usr/local/cuda/lib64/libcublas.so.12"
+            "/usr/local/cuda/lib64/libcublasLt.so.12"
+            "/usr/local/cuda/lib64/libcudart.so.12"
+            "/usr/local/cuda/lib64/libnvToolsExt.so.1"
+            "/usr/local/cuda/lib64/libnvrtc.so.12"
+            "/usr/local/cuda/lib64/libnvrtc-builtins.so"
+        )
+        DEPS_SONAME+=(
+            "libcudnn_adv.so.9"
+            "libcudnn_cnn.so.9"
+            "libcudnn_graph.so.9"
+            "libcudnn_ops.so.9"
+            "libcudnn_engines_runtime_compiled.so.9"
+            "libcudnn_engines_precompiled.so.9"
+            "libcudnn_heuristic.so.9"
+            "libcudnn.so.9"
+            "libcublas.so.12"
+            "libcublasLt.so.12"
+            "libcudart.so.12"
+            "libnvToolsExt.so.1"
+            "libnvrtc.so.12"
+            "libnvrtc-builtins.so"
+        )
+    else
+        echo "Using nvidia libs from pypi."
+        CUDA_RPATHS=(
+            '$ORIGIN/../../nvidia/cublas/lib'
+            '$ORIGIN/../../nvidia/cuda_cupti/lib'
+            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
+            '$ORIGIN/../../nvidia/cuda_runtime/lib'
+            '$ORIGIN/../../nvidia/cudnn/lib'
+            '$ORIGIN/../../nvidia/cufft/lib'
+            '$ORIGIN/../../nvidia/curand/lib'
+            '$ORIGIN/../../nvidia/cusolver/lib'
+            '$ORIGIN/../../nvidia/cusparse/lib'
+            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/nvtx/lib'
+        )
+        CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
+        export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
+        export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
+        export FORCE_RPATH="--force-rpath"
+        export USE_STATIC_NCCL=0
+        export USE_SYSTEM_NCCL=1
+        export ATEN_STATIC_CUDA=0
+        export USE_CUDA_STATIC_LINK=0
+        export USE_CUPTI_SO=1
+        export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+        export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
+    fi
+elif [[ $CUDA_VERSION == "11.8" ]]; then
+    export USE_STATIC_CUDNN=0
+    # Try parallelizing nvcc as well
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+    # Bundle ptxas into the wheel, see https://github.com/pytorch/pytorch/pull/119750
+    export BUILD_BUNDLE_PTXAS=1
+
+    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
+        echo "Bundling with cudnn and cublas."
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcudnn_adv.so.9"
+            "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
+            "/usr/local/cuda/lib64/libcudnn_graph.so.9"
+            "/usr/local/cuda/lib64/libcudnn_ops.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
+            "/usr/local/cuda/lib64/libcudnn.so.9"
+            "/usr/local/cuda/lib64/libcublas.so.11"
+            "/usr/local/cuda/lib64/libcublasLt.so.11"
+            "/usr/local/cuda/lib64/libcudart.so.11.0"
+            "/usr/local/cuda/lib64/libnvToolsExt.so.1"
+            "/usr/local/cuda/lib64/libnvrtc.so.11.2"    # this is not a mistake, it links to more specific cuda version
+            "/usr/local/cuda/lib64/libnvrtc-builtins.so.11.8"
+        )
+        DEPS_SONAME+=(
+            "libcudnn_adv.so.9"
+            "libcudnn_cnn.so.9"
+            "libcudnn_graph.so.9"
+            "libcudnn_ops.so.9"
+            "libcudnn_engines_runtime_compiled.so.9"
+            "libcudnn_engines_precompiled.so.9"
+            "libcudnn_heuristic.so.9"
+            "libcudnn.so.9"
+            "libcublas.so.11"
+            "libcublasLt.so.11"
+            "libcudart.so.11.0"
+            "libnvToolsExt.so.1"
+            "libnvrtc.so.11.2"
+            "libnvrtc-builtins.so.11.8"
+        )
+    else
+        echo "Using nvidia libs from pypi."
+        CUDA_RPATHS=(
+            '$ORIGIN/../../nvidia/cublas/lib'
+            '$ORIGIN/../../nvidia/cuda_cupti/lib'
+            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
+            '$ORIGIN/../../nvidia/cuda_runtime/lib'
+            '$ORIGIN/../../nvidia/cudnn/lib'
+            '$ORIGIN/../../nvidia/cufft/lib'
+            '$ORIGIN/../../nvidia/curand/lib'
+            '$ORIGIN/../../nvidia/cusolver/lib'
+            '$ORIGIN/../../nvidia/cusparse/lib'
+            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/nvtx/lib'
+        )
+        CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
+        export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
+        export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
+        export FORCE_RPATH="--force-rpath"
+        export USE_STATIC_NCCL=0
+        export USE_SYSTEM_NCCL=1
+        export ATEN_STATIC_CUDA=0
+        export USE_CUDA_STATIC_LINK=0
+        export USE_CUPTI_SO=1
+        export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+        export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
+    fi
+else
+    echo "Unknown cuda version $CUDA_VERSION"
+    exit 1
+fi
+
+# builder/test.sh requires DESIRED_CUDA to know what tests to exclude
+export DESIRED_CUDA="$cuda_version_nodot"
+
+# Switch `/usr/local/cuda` to the desired CUDA version
+rm -rf /usr/local/cuda || true
+ln -s "/usr/local/cuda-${CUDA_VERSION}" /usr/local/cuda
+
+# Switch `/usr/local/magma` to the desired CUDA version
+rm -rf /usr/local/magma || true
+ln -s /usr/local/cuda-${CUDA_VERSION}/magma /usr/local/magma
+
+export CUDA_VERSION=$(ls /usr/local/cuda/lib64/libcudart.so.*|sort|tac | head -1 | rev | cut -d"." -f -3 | rev) # 10.0.130
+export CUDA_VERSION_SHORT=$(ls /usr/local/cuda/lib64/libcudart.so.*|sort|tac | head -1 | rev | cut -d"." -f -3 | rev | cut -f1,2 -d".") # 10.0
+export CUDNN_VERSION=$(ls /usr/local/cuda/lib64/libcudnn.so.*|sort|tac | head -1 | rev | cut -d"." -f -3 | rev)
+
+SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P )"
+if [[ -z "$BUILD_PYTHONLESS" ]]; then
+    BUILD_SCRIPT=build_common.sh
+else
+    BUILD_SCRIPT=build_libtorch.sh
+fi
+source $SCRIPTPATH/${BUILD_SCRIPT}
--- a/.ci/manywheel/build_libtorch.sh
+++ b/.ci/manywheel/build_libtorch.sh
@ -0,0 +1,353 @@
+#!/usr/bin/env bash
+# meant to be called only from the neighboring build.sh and build_cpu.sh scripts
+
+set -e pipefail
+SOURCE_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null && pwd )"
+
+# Require only one python installation
+if [[ -z "$DESIRED_PYTHON" ]]; then
+    echo "Need to set DESIRED_PYTHON env variable"
+    exit 1
+fi
+if [[ -n "$BUILD_PYTHONLESS" && -z "$LIBTORCH_VARIANT" ]]; then
+    echo "BUILD_PYTHONLESS is set, so need LIBTORCH_VARIANT to also be set"
+    echo "LIBTORCH_VARIANT should be one of shared-with-deps shared-without-deps static-with-deps static-without-deps"
+    exit 1
+fi
+
+# Function to retry functions that sometimes timeout or have flaky failures
+retry () {
+    $*  || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
+}
+
+# TODO move this into the Docker images
+OS_NAME=`awk -F= '/^NAME/{print $2}' /etc/os-release`
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
+    retry dnf install -q -y zip openssl
+elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
+    # TODO: Remove this once nvidia package repos are back online
+    # Comment out nvidia repositories to prevent them from getting apt-get updated, see https://github.com/pytorch/pytorch/issues/74968
+    # shellcheck disable=SC2046
+    sed -i 's/.*nvidia.*/# &/' $(find /etc/apt/ -type f -name "*.list")
+    retry apt-get update
+    retry apt-get -y install zip openssl
+fi
+
+# Version: setup.py uses $PYTORCH_BUILD_VERSION.post$PYTORCH_BUILD_NUMBER if
+# PYTORCH_BUILD_NUMBER > 1
+build_version="$PYTORCH_BUILD_VERSION"
+build_number="$PYTORCH_BUILD_NUMBER"
+if [[ -n "$OVERRIDE_PACKAGE_VERSION" ]]; then
+    # This will be the *exact* version, since build_number<1
+    build_version="$OVERRIDE_PACKAGE_VERSION"
+    build_number=0
+fi
+if [[ -z "$build_version" ]]; then
+    build_version=1.0.0
+fi
+if [[ -z "$build_number" ]]; then
+    build_number=1
+fi
+export PYTORCH_BUILD_VERSION=$build_version
+export PYTORCH_BUILD_NUMBER=$build_number
+
+export CMAKE_LIBRARY_PATH="/opt/intel/lib:/lib:$CMAKE_LIBRARY_PATH"
+export CMAKE_INCLUDE_PATH="/opt/intel/include:$CMAKE_INCLUDE_PATH"
+
+# set OPENSSL_ROOT_DIR=/opt/openssl if it exists
+if [[ -e /opt/openssl ]]; then
+    export OPENSSL_ROOT_DIR=/opt/openssl
+    export CMAKE_INCLUDE_PATH="/opt/openssl/include":$CMAKE_INCLUDE_PATH
+fi
+
+# If given a python version like 3.6m or 2.7mu, convert this to the format we
+# expect. The binary CI jobs pass in python versions like this; they also only
+# ever pass one python version, so we assume that DESIRED_PYTHON is not a list
+# in this case
+if [[ -n "$DESIRED_PYTHON" && "$DESIRED_PYTHON" != cp* ]]; then
+    python_nodot="$(echo $DESIRED_PYTHON | tr -d m.u)"
+    DESIRED_PYTHON="cp${python_nodot}-cp${python_nodot}"
+fi
+pydir="/opt/python/$DESIRED_PYTHON"
+export PATH="$pydir/bin:$PATH"
+
+export PATCHELF_BIN=/usr/local/bin/patchelf
+patchelf_version=`$PATCHELF_BIN --version`
+echo "patchelf version: " $patchelf_version
+if [[ "$patchelf_version" == "patchelf 0.9" ]]; then
+    echo "Your patchelf version is too old. Please use version >= 0.10."
+    exit 1
+fi
+
+########################################################
+# Compile wheels as well as libtorch
+#######################################################
+if [[ -z "$PYTORCH_ROOT" ]]; then
+    echo "Need to set PYTORCH_ROOT env variable"
+    exit 1
+fi
+pushd "$PYTORCH_ROOT"
+python setup.py clean
+retry pip install -qr requirements.txt
+retry pip install -q numpy==2.0.1
+
+if [[ "$DESIRED_DEVTOOLSET" == *"cxx11-abi"* ]]; then
+    export _GLIBCXX_USE_CXX11_ABI=1
+else
+    export _GLIBCXX_USE_CXX11_ABI=0
+fi
+
+if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+    echo "Calling build_amd.py at $(date)"
+    python tools/amd_build/build_amd.py
+    # TODO remove this work-around once pytorch sources are updated
+    export ROCclr_DIR=/opt/rocm/rocclr/lib/cmake/rocclr
+fi
+
+echo "Calling setup.py install at $(date)"
+
+if [[ $LIBTORCH_VARIANT = *"static"* ]]; then
+    STATIC_CMAKE_FLAG="-DTORCH_STATIC=1"
+fi
+
+(
+    set -x
+
+    mkdir -p build
+
+    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
+        EXTRA_CAFFE2_CMAKE_FLAGS="${EXTRA_CAFFE2_CMAKE_FLAGS[@]} $STATIC_CMAKE_FLAG" \
+        # TODO: Remove this flag once https://github.com/pytorch/pytorch/issues/55952 is closed
+        CFLAGS='-Wno-deprecated-declarations' \
+        BUILD_LIBTORCH_CPU_WITH_DEBUG=1 \
+        python setup.py install
+
+    mkdir -p libtorch/{lib,bin,include,share}
+
+    # Make debug folder separate so it doesn't get zipped up with the rest of
+    # libtorch
+    mkdir debug
+
+    # Copy over all lib files
+    cp -rv build/lib/*                libtorch/lib/
+    cp -rv build/lib*/torch/lib/*     libtorch/lib/
+
+    # Copy over all include files
+    cp -rv build/include/*            libtorch/include/
+    cp -rv build/lib*/torch/include/* libtorch/include/
+
+    # Copy over all of the cmake files
+    cp -rv build/lib*/torch/share/*   libtorch/share/
+
+    # Split libtorch into debug / release version
+    cp libtorch/lib/libtorch_cpu.so libtorch/lib/libtorch_cpu.so.dbg
+
+    # Keep debug symbols on debug lib
+    strip --only-keep-debug libtorch/lib/libtorch_cpu.so.dbg
+
+    # Remove debug info from release lib
+    strip --strip-debug libtorch/lib/libtorch_cpu.so
+
+    # Add a debug link to the release lib to the debug lib (debuggers will then
+    # search for symbols in a file called libtorch_cpu.so.dbg in some
+    # predetermined locations) and embed a CRC32 of the debug library into the .so
+    cd libtorch/lib
+
+    objcopy libtorch_cpu.so --add-gnu-debuglink=libtorch_cpu.so.dbg
+    cd ../..
+
+    # Move the debug symbols to its own directory so it doesn't get processed /
+    # zipped with all the other libraries
+    mv libtorch/lib/libtorch_cpu.so.dbg debug/libtorch_cpu.so.dbg
+
+    echo "${PYTORCH_BUILD_VERSION}" > libtorch/build-version
+    echo "$(pushd $PYTORCH_ROOT && git rev-parse HEAD)" > libtorch/build-hash
+
+)
+
+if [[ "$DESIRED_DEVTOOLSET" == *"cxx11-abi"* ]]; then
+    LIBTORCH_ABI="cxx11-abi-"
+else
+    LIBTORCH_ABI=
+fi
+
+(
+    set -x
+
+    mkdir -p /tmp/$LIBTORCH_HOUSE_DIR
+
+    # objcopy installs a CRC32 into libtorch_cpu above so, so add that to the name here
+    CRC32=$(objcopy --dump-section .gnu_debuglink=>(tail -c4 | od -t x4 -An | xargs echo) libtorch/lib/libtorch_cpu.so)
+
+    # Zip debug symbols
+    zip /tmp/$LIBTORCH_HOUSE_DIR/debug-libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-$PYTORCH_BUILD_VERSION-$CRC32.zip debug/libtorch_cpu.so.dbg
+
+    # Zip and copy libtorch
+    zip -rq /tmp/$LIBTORCH_HOUSE_DIR/libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-$PYTORCH_BUILD_VERSION.zip libtorch
+    cp /tmp/$LIBTORCH_HOUSE_DIR/libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-$PYTORCH_BUILD_VERSION.zip \
+       /tmp/$LIBTORCH_HOUSE_DIR/libtorch-$LIBTORCH_ABI$LIBTORCH_VARIANT-latest.zip
+)
+
+
+popd
+
+#######################################################################
+# ADD DEPENDENCIES INTO THE WHEEL
+#
+# auditwheel repair doesn't work correctly and is buggy
+# so manually do the work of copying dependency libs and patchelfing
+# and fixing RECORDS entries correctly
+######################################################################
+
+fname_with_sha256() {
+    HASH=$(sha256sum $1 | cut -c1-8)
+    DIRNAME=$(dirname $1)
+    BASENAME=$(basename $1)
+    if [[ $BASENAME == "libnvrtc-builtins.so" || $BASENAME == "libcudnn"* ]]; then
+        echo $1
+    else
+        INITNAME=$(echo $BASENAME | cut -f1 -d".")
+        ENDNAME=$(echo $BASENAME | cut -f 2- -d".")
+        echo "$DIRNAME/$INITNAME-$HASH.$ENDNAME"
+    fi
+}
+
+fname_without_so_number() {
+    LINKNAME=$(echo $1 | sed -e 's/\.so.*/.so/g')
+    echo "$LINKNAME"
+}
+
+make_wheel_record() {
+    FPATH=$1
+    if echo $FPATH | grep RECORD >/dev/null 2>&1; then
+        # if the RECORD file, then
+        echo "$FPATH,,"
+    else
+        HASH=$(openssl dgst -sha256 -binary $FPATH | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
+        FSIZE=$(ls -nl $FPATH | awk '{print $5}')
+        echo "$FPATH,sha256=$HASH,$FSIZE"
+    fi
+}
+
+echo 'Built this package:'
+(
+    set -x
+    mkdir -p /$LIBTORCH_HOUSE_DIR
+    mv /tmp/$LIBTORCH_HOUSE_DIR/*.zip /$LIBTORCH_HOUSE_DIR
+    rm -rf /tmp/$LIBTORCH_HOUSE_DIR
+)
+TMP_DIR=$(mktemp -d)
+trap "rm -rf ${TMP_DIR}" EXIT
+pushd "${TMP_DIR}"
+
+for pkg in /$LIBTORCH_HOUSE_DIR/libtorch*.zip; do
+
+    # if the glob didn't match anything
+    if [[ ! -e $pkg ]]; then
+        continue
+    fi
+
+    rm -rf tmp
+    mkdir -p tmp
+    cd tmp
+    cp $pkg .
+
+    unzip -q $(basename $pkg)
+    rm -f $(basename $pkg)
+
+    PREFIX=libtorch
+
+    if [[ $pkg != *"without-deps"* ]]; then
+        # copy over needed dependent .so files over and tag them with their hash
+        patched=()
+        for filepath in "${DEPS_LIST[@]}"; do
+            filename=$(basename $filepath)
+            destpath=$PREFIX/lib/$filename
+            if [[ "$filepath" != "$destpath" ]]; then
+                cp $filepath $destpath
+            fi
+
+            if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+                patchedpath=$(fname_without_so_number $destpath)
+            else
+                patchedpath=$(fname_with_sha256 $destpath)
+            fi
+            patchedname=$(basename $patchedpath)
+            if [[ "$destpath" != "$patchedpath" ]]; then
+                mv $destpath $patchedpath
+            fi
+            patched+=("$patchedname")
+            echo "Copied $filepath to $patchedpath"
+        done
+
+        echo "patching to fix the so names to the hashed names"
+        for ((i=0;i<${#DEPS_LIST[@]};++i)); do
+            find $PREFIX -name '*.so*' | while read sofile; do
+                origname=${DEPS_SONAME[i]}
+                patchedname=${patched[i]}
+                if [[ "$origname" != "$patchedname" ]] || [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
+                    set +e
+                    origname=$($PATCHELF_BIN --print-needed $sofile | grep "$origname.*")
+                    ERRCODE=$?
+                    set -e
+                    if [ "$ERRCODE" -eq "0" ]; then
+                        echo "patching $sofile entry $origname to $patchedname"
+                        $PATCHELF_BIN --replace-needed $origname $patchedname $sofile
+                    fi
+                fi
+            done
+        done
+
+        # copy over needed auxiliary files
+        for ((i=0;i<${#DEPS_AUX_SRCLIST[@]};++i)); do
+            srcpath=${DEPS_AUX_SRCLIST[i]}
+            dstpath=$PREFIX/${DEPS_AUX_DSTLIST[i]}
+            mkdir -p $(dirname $dstpath)
+            cp $srcpath $dstpath
+        done
+    fi
+
+    # set RPATH of _C.so and similar to $ORIGIN, $ORIGIN/lib
+    find $PREFIX -maxdepth 1 -type f -name "*.so*" | while read sofile; do
+        echo "Setting rpath of $sofile to " '$ORIGIN:$ORIGIN/lib'
+        $PATCHELF_BIN --set-rpath '$ORIGIN:$ORIGIN/lib' $sofile
+        $PATCHELF_BIN --print-rpath $sofile
+    done
+
+    # set RPATH of lib/ files to $ORIGIN
+    find $PREFIX/lib -maxdepth 1 -type f -name "*.so*" | while read sofile; do
+        echo "Setting rpath of $sofile to " '$ORIGIN'
+        $PATCHELF_BIN --set-rpath '$ORIGIN' $sofile
+        $PATCHELF_BIN --print-rpath $sofile
+    done
+
+    # regenerate the RECORD file with new hashes
+    record_file=`echo $(basename $pkg) | sed -e 's/-cp.*$/.dist-info\/RECORD/g'`
+    if [[ -e $record_file ]]; then
+        echo "Generating new record file $record_file"
+        rm -f $record_file
+        # generate records for folders in wheel
+        find * -type f | while read fname; do
+            echo $(make_wheel_record $fname) >>$record_file
+        done
+    fi
+
+    # zip up the wheel back
+    zip -rq $(basename $pkg) $PREFIX*
+
+    # replace original wheel
+    rm -f $pkg
+    mv $(basename $pkg) $pkg
+    cd ..
+    rm -rf tmp
+done
+
+# Copy wheels to host machine for persistence before testing
+if [[ -n "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
+    cp /$LIBTORCH_HOUSE_DIR/libtorch*.zip "$PYTORCH_FINAL_PACKAGE_DIR"
+    cp /$LIBTORCH_HOUSE_DIR/debug-libtorch*.zip "$PYTORCH_FINAL_PACKAGE_DIR"
+fi
--- a/.ci/manywheel/build_rocm.sh
+++ b/.ci/manywheel/build_rocm.sh
@ -0,0 +1,263 @@
+#!/usr/bin/env bash
+
+set -ex
+
+export ROCM_HOME=/opt/rocm
+export MAGMA_HOME=$ROCM_HOME/magma
+# TODO: libtorch_cpu.so is broken when building with Debug info
+export BUILD_DEBUG_INFO=0
+
+# TODO Are these all used/needed?
+export TH_BINARY_BUILD=1
+export USE_STATIC_CUDNN=1
+export USE_STATIC_NCCL=1
+export ATEN_STATIC_CUDA=1
+export USE_CUDA_STATIC_LINK=1
+export INSTALL_TEST=0 # dont install test binaries into site-packages
+# Set RPATH instead of RUNPATH when using patchelf to avoid LD_LIBRARY_PATH override
+export FORCE_RPATH="--force-rpath"
+
+# Keep an array of cmake variables to add to
+if [[ -z "$CMAKE_ARGS" ]]; then
+    # These are passed to tools/build_pytorch_libs.sh::build()
+    CMAKE_ARGS=()
+fi
+if [[ -z "$EXTRA_CAFFE2_CMAKE_FLAGS" ]]; then
+    # These are passed to tools/build_pytorch_libs.sh::build_caffe2()
+    EXTRA_CAFFE2_CMAKE_FLAGS=()
+fi
+
+# Determine ROCm version and architectures to build for
+#
+# NOTE: We should first check `DESIRED_CUDA` when determining `ROCM_VERSION`
+if [[ -n "$DESIRED_CUDA" ]]; then
+    if ! echo "${DESIRED_CUDA}"| grep "^rocm" >/dev/null 2>/dev/null; then
+        export DESIRED_CUDA="rocm${DESIRED_CUDA}"
+    fi
+    # rocm3.7, rocm3.5.1
+    ROCM_VERSION="$DESIRED_CUDA"
+    echo "Using $ROCM_VERSION as determined by DESIRED_CUDA"
+else
+    echo "Must set DESIRED_CUDA"
+    exit 1
+fi
+
+# Package directories
+WHEELHOUSE_DIR="wheelhouse$ROCM_VERSION"
+LIBTORCH_HOUSE_DIR="libtorch_house$ROCM_VERSION"
+if [[ -z "$PYTORCH_FINAL_PACKAGE_DIR" ]]; then
+    if [[ -z "$BUILD_PYTHONLESS" ]]; then
+        PYTORCH_FINAL_PACKAGE_DIR="/remote/wheelhouse$ROCM_VERSION"
+    else
+        PYTORCH_FINAL_PACKAGE_DIR="/remote/libtorch_house$ROCM_VERSION"
+    fi
+fi
+mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
+
+# To make version comparison easier, create an integer representation.
+ROCM_VERSION_CLEAN=$(echo ${ROCM_VERSION} | sed s/rocm//)
+save_IFS="$IFS"
+IFS=. ROCM_VERSION_ARRAY=(${ROCM_VERSION_CLEAN})
+IFS="$save_IFS"
+if [[ ${#ROCM_VERSION_ARRAY[@]} == 2 ]]; then
+    ROCM_VERSION_MAJOR=${ROCM_VERSION_ARRAY[0]}
+    ROCM_VERSION_MINOR=${ROCM_VERSION_ARRAY[1]}
+    ROCM_VERSION_PATCH=0
+elif [[ ${#ROCM_VERSION_ARRAY[@]} == 3 ]]; then
+    ROCM_VERSION_MAJOR=${ROCM_VERSION_ARRAY[0]}
+    ROCM_VERSION_MINOR=${ROCM_VERSION_ARRAY[1]}
+    ROCM_VERSION_PATCH=${ROCM_VERSION_ARRAY[2]}
+else
+    echo "Unhandled ROCM_VERSION ${ROCM_VERSION}"
+    exit 1
+fi
+ROCM_INT=$(($ROCM_VERSION_MAJOR * 10000 + $ROCM_VERSION_MINOR * 100 + $ROCM_VERSION_PATCH))
+
+# Required ROCm libraries
+ROCM_SO_FILES=(
+    "libMIOpen.so"
+    "libamdhip64.so"
+    "libhipblas.so"
+    "libhipfft.so"
+    "libhiprand.so"
+    "libhipsolver.so"
+    "libhipsparse.so"
+    "libhsa-runtime64.so"
+    "libamd_comgr.so"
+    "libmagma.so"
+    "librccl.so"
+    "librocblas.so"
+    "librocfft.so"
+    "librocm_smi64.so"
+    "librocrand.so"
+    "librocsolver.so"
+    "librocsparse.so"
+    "libroctracer64.so"
+    "libroctx64.so"
+    "libhipblaslt.so"
+    "libhiprtc.so"
+)
+
+if [[ $ROCM_INT -ge 60100 ]]; then
+    ROCM_SO_FILES+=("librocprofiler-register.so")
+fi
+
+if [[ $ROCM_INT -ge 60200 ]]; then
+    ROCM_SO_FILES+=("librocm-core.so")
+fi
+
+OS_NAME=`awk -F= '/^NAME/{print $2}' /etc/os-release`
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+    LIBNUMA_PATH="/usr/lib64/libnuma.so.1"
+    LIBELF_PATH="/usr/lib64/libelf.so.1"
+    LIBTINFO_PATH="/usr/lib64/libtinfo.so.5"
+    LIBDRM_PATH="/opt/amdgpu/lib64/libdrm.so.2"
+    LIBDRM_AMDGPU_PATH="/opt/amdgpu/lib64/libdrm_amdgpu.so.1"
+    if [[ $ROCM_INT -ge 60100 ]]; then
+        # Below libs are direct dependencies of libhipsolver
+        LIBSUITESPARSE_CONFIG_PATH="/lib64/libsuitesparseconfig.so.4"
+        LIBCHOLMOD_PATH="/lib64/libcholmod.so.2"
+        # Below libs are direct dependencies of libcholmod
+        LIBAMD_PATH="/lib64/libamd.so.2"
+        LIBCAMD_PATH="/lib64/libcamd.so.2"
+        LIBCCOLAMD_PATH="/lib64/libccolamd.so.2"
+        LIBCOLAMD_PATH="/lib64/libcolamd.so.2"
+        LIBSATLAS_PATH="/lib64/atlas/libsatlas.so.3"
+        # Below libs are direct dependencies of libsatlas
+        LIBGFORTRAN_PATH="/lib64/libgfortran.so.3"
+        LIBQUADMATH_PATH="/lib64/libquadmath.so.0"
+    fi
+    MAYBE_LIB64=lib64
+elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
+    LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
+    LIBNUMA_PATH="/usr/lib/x86_64-linux-gnu/libnuma.so.1"
+    LIBELF_PATH="/usr/lib/x86_64-linux-gnu/libelf.so.1"
+    if [[ $ROCM_INT -ge 50300 ]]; then
+        LIBTINFO_PATH="/lib/x86_64-linux-gnu/libtinfo.so.6"
+    else
+        LIBTINFO_PATH="/lib/x86_64-linux-gnu/libtinfo.so.5"
+    fi
+    LIBDRM_PATH="/usr/lib/x86_64-linux-gnu/libdrm.so.2"
+    LIBDRM_AMDGPU_PATH="/usr/lib/x86_64-linux-gnu/libdrm_amdgpu.so.1"
+    if [[ $ROCM_INT -ge 60100 ]]; then
+        # Below libs are direct dependencies of libhipsolver
+        LIBCHOLMOD_PATH="/lib/x86_64-linux-gnu/libcholmod.so.3"
+        # Below libs are direct dependencies of libcholmod
+        LIBSUITESPARSE_CONFIG_PATH="/lib/x86_64-linux-gnu/libsuitesparseconfig.so.5"
+        LIBAMD_PATH="/lib/x86_64-linux-gnu/libamd.so.2"
+        LIBCAMD_PATH="/lib/x86_64-linux-gnu/libcamd.so.2"
+        LIBCCOLAMD_PATH="/lib/x86_64-linux-gnu/libccolamd.so.2"
+        LIBCOLAMD_PATH="/lib/x86_64-linux-gnu/libcolamd.so.2"
+        LIBMETIS_PATH="/lib/x86_64-linux-gnu/libmetis.so.5"
+        LIBLAPACK_PATH="/lib/x86_64-linux-gnu/liblapack.so.3"
+        LIBBLAS_PATH="/lib/x86_64-linux-gnu/libblas.so.3"
+        # Below libs are direct dependencies of libblas
+        LIBGFORTRAN_PATH="/lib/x86_64-linux-gnu/libgfortran.so.5"
+        LIBQUADMATH_PATH="/lib/x86_64-linux-gnu/libquadmath.so.0"
+    fi
+    MAYBE_LIB64=lib
+fi
+OS_SO_PATHS=($LIBGOMP_PATH $LIBNUMA_PATH\
+             $LIBELF_PATH $LIBTINFO_PATH\
+             $LIBDRM_PATH $LIBDRM_AMDGPU_PATH\
+             $LIBSUITESPARSE_CONFIG_PATH\
+             $LIBCHOLMOD_PATH $LIBAMD_PATH\
+             $LIBCAMD_PATH $LIBCCOLAMD_PATH\
+             $LIBCOLAMD_PATH $LIBSATLAS_PATH\
+             $LIBGFORTRAN_PATH $LIBQUADMATH_PATH\
+             $LIBMETIS_PATH $LIBLAPACK_PATH\
+             $LIBBLAS_PATH)
+OS_SO_FILES=()
+for lib in "${OS_SO_PATHS[@]}"
+do
+    file_name="${lib##*/}" # Substring removal of path to get filename
+    OS_SO_FILES[${#OS_SO_FILES[@]}]=$file_name # Append lib to array
+done
+
+# PyTorch-version specific
+# AOTriton dependency only for PyTorch >= 2.4
+if (( $(echo "${PYTORCH_VERSION} 2.4" | awk '{print ($1 >= $2)}') )); then
+    ROCM_SO_FILES+=("libaotriton_v2.so")
+fi
+
+# rocBLAS library files
+ROCBLAS_LIB_SRC=$ROCM_HOME/lib/rocblas/library
+ROCBLAS_LIB_DST=lib/rocblas/library
+ARCH=$(echo $PYTORCH_ROCM_ARCH | sed 's/;/|/g') # Replace ; seperated arch list to bar for grep
+ARCH_SPECIFIC_FILES=$(ls $ROCBLAS_LIB_SRC | grep -E $ARCH)
+OTHER_FILES=$(ls $ROCBLAS_LIB_SRC | grep -v gfx)
+ROCBLAS_LIB_FILES=($ARCH_SPECIFIC_FILES $OTHER_FILES)
+
+# hipblaslt library files
+HIPBLASLT_LIB_SRC=$ROCM_HOME/lib/hipblaslt/library
+HIPBLASLT_LIB_DST=lib/hipblaslt/library
+ARCH_SPECIFIC_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -E $ARCH)
+OTHER_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -v gfx)
+HIPBLASLT_LIB_FILES=($ARCH_SPECIFIC_FILES $OTHER_FILES)
+
+# ROCm library files
+ROCM_SO_PATHS=()
+for lib in "${ROCM_SO_FILES[@]}"
+do
+    file_path=($(find $ROCM_HOME/lib/ -name "$lib")) # First search in lib
+    if [[ -z $file_path ]]; then
+        if [ -d "$ROCM_HOME/lib64/" ]; then
+            file_path=($(find $ROCM_HOME/lib64/ -name "$lib")) # Then search in lib64
+        fi
+    fi
+    if [[ -z $file_path ]]; then
+        file_path=($(find $ROCM_HOME/ -name "$lib")) # Then search in ROCM_HOME
+    fi
+    if [[ -z $file_path ]]; then
+        echo "Error: Library file $lib is not found." >&2
+        exit 1
+    fi
+    ROCM_SO_PATHS[${#ROCM_SO_PATHS[@]}]="$file_path" # Append lib to array
+done
+
+DEPS_LIST=(
+    ${ROCM_SO_PATHS[*]}
+    ${OS_SO_PATHS[*]}
+)
+
+DEPS_SONAME=(
+    ${ROCM_SO_FILES[*]}
+    ${OS_SO_FILES[*]}
+)
+
+DEPS_AUX_SRCLIST=(
+    "${ROCBLAS_LIB_FILES[@]/#/$ROCBLAS_LIB_SRC/}"
+    "${HIPBLASLT_LIB_FILES[@]/#/$HIPBLASLT_LIB_SRC/}"
+    "/opt/amdgpu/share/libdrm/amdgpu.ids"
+)
+
+DEPS_AUX_DSTLIST=(
+    "${ROCBLAS_LIB_FILES[@]/#/$ROCBLAS_LIB_DST/}"
+    "${HIPBLASLT_LIB_FILES[@]/#/$HIPBLASLT_LIB_DST/}"
+    "share/libdrm/amdgpu.ids"
+)
+
+# MIOpen library files
+MIOPEN_SHARE_SRC=$ROCM_HOME/share/miopen/db
+MIOPEN_SHARE_DST=share/miopen/db
+MIOPEN_SHARE_FILES=($(ls $MIOPEN_SHARE_SRC | grep -E $ARCH))
+DEPS_AUX_SRCLIST+=(${MIOPEN_SHARE_FILES[@]/#/$MIOPEN_SHARE_SRC/})
+DEPS_AUX_DSTLIST+=(${MIOPEN_SHARE_FILES[@]/#/$MIOPEN_SHARE_DST/})
+
+# RCCL library files
+RCCL_SHARE_SRC=$ROCM_HOME/share/rccl/msccl-algorithms
+RCCL_SHARE_DST=share/rccl/msccl-algorithms
+RCCL_SHARE_FILES=($(ls $RCCL_SHARE_SRC))
+DEPS_AUX_SRCLIST+=(${RCCL_SHARE_FILES[@]/#/$RCCL_SHARE_SRC/})
+DEPS_AUX_DSTLIST+=(${RCCL_SHARE_FILES[@]/#/$RCCL_SHARE_DST/})
+
+echo "PYTORCH_ROCM_ARCH: ${PYTORCH_ROCM_ARCH}"
+
+SCRIPTPATH="$( cd "$(dirname "$0")" ; pwd -P )"
+if [[ -z "$BUILD_PYTHONLESS" ]]; then
+    BUILD_SCRIPT=build_common.sh
+else
+    BUILD_SCRIPT=build_libtorch.sh
+fi
+source $SCRIPTPATH/${BUILD_SCRIPT}
--- a/.ci/manywheel/test_wheel.sh
+++ b/.ci/manywheel/test_wheel.sh
@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+set -e
+
+yum install -y wget git
+
+rm -rf /usr/local/cuda*
+
+# Install Anaconda
+if ! ls /py
+then
+    echo "Miniconda needs to be installed"
+    wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda.sh
+    bash ~/miniconda.sh -b -p /py
+else
+    echo "Miniconda is already installed"
+fi
+
+export PATH="/py/bin:$PATH"
+
+# Anaconda token
+if ls /remote/token
+then
+   source /remote/token
+fi
+
+conda install -y conda-build anaconda-client
--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -205,7 +205,8 @@ fi
 if [[ "$BUILD_ENVIRONMENT" == *-clang*-asan* ]]; then
  export USE_CUDA=0
  export USE_ASAN=1
-  export UBSAN_FLAGS="-fno-sanitize-recover=all;-fno-sanitize=float-divide-by-zero;-fno-sanitize=float-cast-overflow"
+  export REL_WITH_DEB_INFO=1
+  export UBSAN_FLAGS="-fno-sanitize-recover=all"
  unset USE_LLVM
 fi

@ -273,7 +274,6 @@ else
    # set only when building other architectures
    # or building non-XLA tests.
    if [[ "$BUILD_ENVIRONMENT" != *rocm*  &&
-          "$BUILD_ENVIRONMENT" != *s390x*   &&
          "$BUILD_ENVIRONMENT" != *xla* ]]; then
      if [[ "$BUILD_ENVIRONMENT" != *py3.8* ]]; then
        # Install numpy-2.0.2 for builds which are backward compatible with 1.X
--- a/.ci/pytorch/test.sh
+++ b/.ci/pytorch/test.sh
@ -320,7 +320,6 @@ test_inductor_distributed() {
  python test/run_test.py -i distributed/test_c10d_functional_native.py --verbose
  python test/run_test.py -i distributed/_tensor/test_dtensor_compile.py --verbose
  python test/run_test.py -i distributed/tensor/parallel/test_micro_pipeline_tp.py --verbose
-  python test/run_test.py -i distributed/_composable/test_replicate_with_compiler.py --verbose
  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_comm.py --verbose
  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_training.py -k test_train_parity_multi_group --verbose
  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_training.py -k test_train_parity_with_activation_checkpointing --verbose
@ -332,7 +331,6 @@ test_inductor_distributed() {
  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_mixed_precision.py -k test_compute_dtype --verbose
  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_mixed_precision.py -k test_reduce_dtype --verbose
  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_clip_grad_norm_.py -k test_clip_grad_norm_2d --verbose
-  python test/run_test.py -i distributed/_composable/fsdp/test_fully_shard_compile.py --verbose
  python test/run_test.py -i distributed/fsdp/test_fsdp_tp_integration.py -k test_fsdp_tp_integration --verbose

  # this runs on both single-gpu and multi-gpu instance. It should be smart about skipping tests that aren't supported
--- a/.ci/pytorch/win-test.sh
+++ b/.ci/pytorch/win-test.sh
@ -46,6 +46,9 @@ python -m pip install tlparse==0.3.25
 # Install parameterized
 python -m pip install parameterized==0.8.1

+# Install pulp for testing ilps under torch\distributed\_tools
+python -m pip install pulp==2.9.0
+
 run_tests() {
    # Run nvidia-smi if available
    for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do
--- a/.github/actions/setup-linux/action.yml
+++ b/.github/actions/setup-linux/action.yml
@ -20,7 +20,7 @@ runs:
          elif [[ $runner_name_str == *"gcp"* ]]; then
            echo "Runner is from Google Cloud Platform, No info on ec2 metadata"
          else
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          fi
        }
        echo "ami-id: $(get_ec2_metadata ami-id)"
--- a/.github/actions/setup-win/action.yml
+++ b/.github/actions/setup-win/action.yml
@ -18,7 +18,7 @@ runs:
          # Pulled from instance metadata endpoint for EC2
          # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
          category=$1
-          curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
        }
        echo "ami-id: $(get_ec2_metadata ami-id)"
        echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/ci_commit_pins/audio.txt
+++ b/.github/ci_commit_pins/audio.txt
@ -1 +1 @@
-3f0569939c4369bec943fc27d1c9d8dfbc828c26
+79047bf6bdec9e32c4cffd0f9835b347781fefbf
--- a/.github/lf-canary-scale-config.yml
+++ b/.github/lf-canary-scale-config.yml
@ -1,251 +0,0 @@
-
-# This file is generated by .github/scripts/validate_scale_config.py in test-infra
-# It defines runner types that will be provisioned by by LF Self-hosted runners
-
-# scale-config.yml:
-#   Powers what instance types are available for GHA auto-scaled
-#   runners. Runners listed here will be available as self hosted
-#   runners, configuration is directly pulled from the main branch.
-#
-#
-# NOTES:
-#  - Linux runners are by default non-ephemeral to reduce the amount of CreateInstaces calls
-#    to avoid RequestLimitExceeded issues
-#  - When updating this file, run the following command to validate the YAML and to generate
-#    corresponding versions of scale-config for the pytorch/pytorch repo and merge the
-#    pytorch/pytorch changes before merging these changes.
-#    `python .github/scripts/validate_scale_config.py --test-infra-repo-root [path_to_test-infra_root] --pytorch-repo-root [path_to_pytorch_root]``
-#
-# TODO: Add some documentation on how the auto-scaling works
-#
-# NOTE: Default values,
-#
-# runner_types:
-#   runner_label:
-#     instance_type: m4.large
-#     os: linux
-#     max_available: 20
-#     disk_size: 50
-#     is_ephemeral: true
-
-runner_types:
-  lf.c.linux.12xlarge:
-    disk_size: 200
-    instance_type: c5.12xlarge
-    is_ephemeral: false
-    max_available: 1000
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.10xlarge.avx2:
-    disk_size: 200
-    instance_type: m4.10xlarge
-    is_ephemeral: false
-    max_available: 450
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.24xl.spr-metal:
-    disk_size: 200
-    instance_type: c7i.metal-24xl
-    is_ephemeral: false
-    max_available: 150
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.16xlarge.spr:
-    disk_size: 200
-    instance_type: c7i.16xlarge
-    is_ephemeral: false
-    max_available: 150
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.9xlarge.ephemeral:
-    disk_size: 200
-    instance_type: c5.9xlarge
-    is_ephemeral: true
-    max_available: 50
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-    variants:
-      am2:
-        ami: amzn2-ami-hvm-2.0.20240306.2-x86_64-ebs
-  lf.c.linux.12xlarge.ephemeral:
-    disk_size: 200
-    instance_type: c5.12xlarge
-    is_ephemeral: true
-    max_available: 300
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.16xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g3.16xlarge
-    is_ephemeral: false
-    max_available: 150
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.24xlarge:
-    disk_size: 150
-    instance_type: c5.24xlarge
-    is_ephemeral: false
-    max_available: 500
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.24xlarge.ephemeral:
-    disk_size: 150
-    instance_type: c5.24xlarge
-    is_ephemeral: true
-    max_available: 200
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.2xlarge:
-    disk_size: 150
-    instance_type: c5.2xlarge
-    is_ephemeral: false
-    max_available: 3120
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.4xlarge:
-    disk_size: 150
-    instance_type: c5.4xlarge
-    is_ephemeral: false
-    max_available: 1000
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.4xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g3.4xlarge
-    is_ephemeral: false
-    max_available: 1000
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.8xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g3.8xlarge
-    is_ephemeral: false
-    max_available: 400
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.g4dn.12xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g4dn.12xlarge
-    is_ephemeral: false
-    max_available: 250
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.g4dn.metal.nvidia.gpu:
-    disk_size: 150
-    instance_type: g4dn.metal
-    is_ephemeral: false
-    max_available: 300
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.g5.48xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g5.48xlarge
-    is_ephemeral: false
-    max_available: 200
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.g5.12xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g5.12xlarge
-    is_ephemeral: false
-    max_available: 150
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.g5.4xlarge.nvidia.gpu:
-    disk_size: 150
-    instance_type: g5.4xlarge
-    is_ephemeral: false
-    max_available: 2400
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.g6.4xlarge.experimental.nvidia.gpu:
-    disk_size: 150
-    instance_type: g6.4xlarge
-    is_ephemeral: false
-    max_available: 50
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.large:
-    max_available: 1200
-    disk_size: 15
-    instance_type: c5.large
-    is_ephemeral: false
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
-  lf.c.linux.arm64.2xlarge:
-    disk_size: 256
-    instance_type: t4g.2xlarge
-    is_ephemeral: false
-    max_available: 200
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-arm64
-  lf.c.linux.arm64.m7g.4xlarge:
-    disk_size: 256
-    instance_type: m7g.4xlarge
-    is_ephemeral: false
-    max_available: 200
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-arm64
-  lf.c.linux.arm64.2xlarge.ephemeral:
-    disk_size: 256
-    instance_type: t4g.2xlarge
-    is_ephemeral: true
-    max_available: 200
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-arm64
-  lf.c.linux.arm64.m7g.4xlarge.ephemeral:
-    disk_size: 256
-    instance_type: m7g.4xlarge
-    is_ephemeral: true
-    max_available: 200
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-arm64
-  lf.c.linux.arm64.m7g.metal:
-    disk_size: 256
-    instance_type: m7g.metal
-    is_ephemeral: false
-    max_available: 100
-    os: linux
-    ami: al2023-ami-2023.5.202*-kernel-6.1-arm64
-  lf.c.windows.g4dn.xlarge:
-    disk_size: 256
-    instance_type: g4dn.xlarge
-    is_ephemeral: true
-    max_available: 100
-    os: windows
-  lf.c.windows.g4dn.xlarge.nonephemeral:
-    disk_size: 256
-    instance_type: g4dn.xlarge
-    is_ephemeral: false
-    max_available: 100
-    os: windows
-  lf.c.windows.4xlarge:
-    disk_size: 256
-    instance_type: c5d.4xlarge
-    is_ephemeral: true
-    max_available: 420
-    os: windows
-  lf.c.windows.4xlarge.nonephemeral:
-    disk_size: 256
-    instance_type: c5d.4xlarge
-    is_ephemeral: false
-    max_available: 420
-    os: windows
-  lf.c.windows.8xlarge.nvidia.gpu:
-    disk_size: 256
-    instance_type: p3.2xlarge
-    is_ephemeral: true
-    max_available: 300
-    os: windows
-  lf.c.windows.8xlarge.nvidia.gpu.nonephemeral:
-    disk_size: 256
-    instance_type: p3.2xlarge
-    is_ephemeral: false
-    max_available: 150
-    os: windows
-  lf.c.windows.g5.4xlarge.nvidia.gpu:
-    disk_size: 256
-    instance_type: g5.4xlarge
-    is_ephemeral: false
-    max_available: 250
-    os: windows
--- a/.github/lf-scale-config.yml
+++ b/.github/lf-scale-config.yml
@ -33,7 +33,7 @@ runner_types:
    disk_size: 200
    instance_type: c5.12xlarge
    is_ephemeral: false
-    max_available: 1000
+    max_available: 2000
    os: linux
    ami: al2023-ami-2023.5.202*-kernel-6.1-x86_64
  lf.linux.10xlarge.avx2:
@ -241,7 +241,7 @@ runner_types:
    disk_size: 256
    instance_type: p3.2xlarge
    is_ephemeral: false
-    max_available: 150
+    max_available: 300
    os: windows
  lf.windows.g5.4xlarge.nvidia.gpu:
    disk_size: 256
--- a/.github/pytorch-probot.yml
+++ b/.github/pytorch-probot.yml
@ -22,6 +22,7 @@ ciflow_push_tags:
 - ciflow/unstable
 - ciflow/xpu
 - ciflow/torchbench
+- ciflow/autoformat
 retryable_workflows:
 - pull
 - trunk
--- a/.github/scripts/lintrunner.sh
+++ b/.github/scripts/lintrunner.sh
@ -41,7 +41,8 @@ RC=0
 if ! lintrunner --force-color --tee-json=lint.json ${ADDITIONAL_LINTRUNNER_ARGS} 2> /dev/null; then
    echo ""
    echo -e "\e[1m\e[36mYou can reproduce these results locally by using \`lintrunner -m origin/main\`. (If you don't get the same results, run \'lintrunner init\' to update your local linter)\e[0m"
-    echo -e "\e[1m\e[36mSee https://github.com/pytorch/pytorch/wiki/lintrunner for setup instructions.\e[0m"
+    echo -e "\e[1m\e[36mSee https://github.com/pytorch/pytorch/wiki/lintrunner for setup instructions. To apply suggested patches automatically, use the -a flag. Before pushing another commit,\e[0m"
+    echo -e "\e[1m\e[36mplease verify locally and ensure everything passes.\e[0m"
    RC=1
 fi

--- a/.github/templates/common.yml.j2
+++ b/.github/templates/common.yml.j2
@ -25,7 +25,7 @@ concurrency:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/_binary-build-linux.yml
+++ b/.github/workflows/_binary-build-linux.yml
@ -272,6 +272,8 @@ jobs:
          docker exec -t -w "${PYTORCH_ROOT}" "${container_name}" bash -c "bash .circleci/scripts/binary_populate_env.sh"
          if [[ ${BUILD_ENVIRONMENT} == *"aarch64"* ]]; then
            docker exec -t "${container_name}" bash -c "bash /builder/aarch64_linux/aarch64_ci_build.sh"
+          elif [[ ${{ inputs.PACKAGE_TYPE }} == "manywheel" || ${{ inputs.PACKAGE_TYPE }} == "libtorch" ]]; then
+            docker exec -t "${container_name}" bash -c "source ${BINARY_ENV_FILE} && bash /pytorch/.ci/${{ inputs.PACKAGE_TYPE }}/build.sh"
          else
            docker exec -t "${container_name}" bash -c "source ${BINARY_ENV_FILE} && bash /builder/${{ inputs.PACKAGE_TYPE }}/build.sh"
          fi
--- a/.github/workflows/generated-windows-binary-conda-nightly.yml
+++ b/.github/workflows/generated-windows-binary-conda-nightly.yml
@ -64,7 +64,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -178,7 +178,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -310,7 +310,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -425,7 +425,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -558,7 +558,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -673,7 +673,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -806,7 +806,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -921,7 +921,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1053,7 +1053,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1167,7 +1167,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1299,7 +1299,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1414,7 +1414,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1547,7 +1547,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1662,7 +1662,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1795,7 +1795,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1910,7 +1910,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2042,7 +2042,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2156,7 +2156,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2288,7 +2288,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2403,7 +2403,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2536,7 +2536,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2651,7 +2651,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2784,7 +2784,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2899,7 +2899,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3031,7 +3031,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3145,7 +3145,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3277,7 +3277,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3392,7 +3392,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3525,7 +3525,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3640,7 +3640,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3773,7 +3773,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3888,7 +3888,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/generated-windows-binary-libtorch-debug-main.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-debug-main.yml
@ -61,7 +61,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -179,7 +179,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml
@ -68,7 +68,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -186,7 +186,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -326,7 +326,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -445,7 +445,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -586,7 +586,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -705,7 +705,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -846,7 +846,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -965,7 +965,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/generated-windows-binary-libtorch-release-main.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-release-main.yml
@ -61,7 +61,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -179,7 +179,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/generated-windows-binary-libtorch-release-nightly.yml
+++ b/.github/workflows/generated-windows-binary-libtorch-release-nightly.yml
@ -68,7 +68,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -186,7 +186,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -326,7 +326,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -445,7 +445,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -586,7 +586,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -705,7 +705,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -846,7 +846,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -965,7 +965,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/generated-windows-binary-wheel-nightly.yml
+++ b/.github/workflows/generated-windows-binary-wheel-nightly.yml
@ -65,7 +65,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -179,7 +179,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -312,7 +312,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -427,7 +427,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -561,7 +561,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -676,7 +676,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -810,7 +810,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -925,7 +925,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1057,7 +1057,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1171,7 +1171,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1303,7 +1303,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1417,7 +1417,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1550,7 +1550,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1665,7 +1665,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1799,7 +1799,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -1914,7 +1914,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2048,7 +2048,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2163,7 +2163,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2295,7 +2295,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2409,7 +2409,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2541,7 +2541,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2655,7 +2655,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2788,7 +2788,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -2903,7 +2903,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3037,7 +3037,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3152,7 +3152,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3286,7 +3286,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3401,7 +3401,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3533,7 +3533,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3647,7 +3647,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3779,7 +3779,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -3893,7 +3893,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4026,7 +4026,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4141,7 +4141,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4275,7 +4275,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4390,7 +4390,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4524,7 +4524,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4639,7 +4639,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4771,7 +4771,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
@ -4885,7 +4885,7 @@ jobs:
            # Pulled from instance metadata endpoint for EC2
            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
            category=$1
-            curl -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
          }
          echo "ami-id: $(get_ec2_metadata ami-id)"
          echo "instance-id: $(get_ec2_metadata instance-id)"
--- a/.github/workflows/lint-autoformat.yml
+++ b/.github/workflows/lint-autoformat.yml
@ -2,9 +2,9 @@ name: Apply lint suggestions

 on:

-  pull_request:
-    types: [opened, synchronize, reopened]
-    branches: [main]
+  push:
+    tags:
+      - ciflow/autoformat/*

 jobs:
  lintrunner-autoformat:
--- a/.github/workflows/update-viablestrict.yml
+++ b/.github/workflows/update-viablestrict.yml
@ -25,7 +25,9 @@ jobs:
          stable-branch: viable/strict
          requires: '[\"pull\", \"trunk\", \"lint\", \"linux-binary\"]'
          secret-bot-token: ${{ secrets.MERGEBOT_TOKEN }}
-          rockset-api-key: ${{ secrets.ROCKSET_API_KEY }}
+          clickhouse-url: ${{ secrets.CLICKHOUSE_URL }}
+          clickhouse-username: ${{ secrets.CLICKHOUSE_VIABLESTRICT_USERNAME }}
+          clickhouse-password: ${{ secrets.CLICKHOUSE_VIABLESTRICT_PASSWORD }}

      - name: Authenticate to AWS with OIDC
        uses: aws-actions/configure-aws-credentials@v4
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -341,19 +341,6 @@ cmake_dependent_option(
 cmake_dependent_option(USE_SYSTEM_UCC "Use system-wide UCC" OFF "USE_UCC" OFF)
 cmake_dependent_option(USE_C10D_UCC "USE C10D UCC" ON "USE_DISTRIBUTED;USE_UCC"
                       OFF)
-cmake_dependent_option(
-  USE_GLOO "Use Gloo. Only available if USE_DISTRIBUTED is on." ON
-  "USE_DISTRIBUTED" OFF)
-cmake_dependent_option(
-  USE_GLOO_WITH_OPENSSL
-  "Use Gloo with OpenSSL. Only available if USE_GLOO is on." OFF
-  "USE_GLOO AND LINUX AND NOT INTERN_BUILD_MOBILE" OFF)
-cmake_dependent_option(USE_C10D_GLOO "USE C10D GLOO" ON
-                       "USE_DISTRIBUTED;USE_GLOO" OFF)
-cmake_dependent_option(USE_C10D_NCCL "USE C10D NCCL" ON
-                       "USE_DISTRIBUTED;USE_NCCL" OFF)
-cmake_dependent_option(USE_C10D_MPI "USE C10D MPI" ON "USE_DISTRIBUTED;USE_MPI"
-                       OFF)
 cmake_dependent_option(
    USE_GLOO "Use Gloo. Only available if USE_DISTRIBUTED is on." ON
    "USE_DISTRIBUTED" OFF)
@ -469,6 +456,7 @@ option(USE_SYSTEM_FXDIV "Use system-provided fxdiv." OFF)
 option(USE_SYSTEM_BENCHMARK "Use system-provided google benchmark." OFF)
 option(USE_SYSTEM_ONNX "Use system-provided onnx." OFF)
 option(USE_SYSTEM_XNNPACK "Use system-provided xnnpack." OFF)
+OPTION(USE_SYSTEM_NVTX "Use system-provided nvtx." OFF)
 option(USE_GOLD_LINKER "Use ld.gold to link" OFF)
 if(USE_SYSTEM_LIBS)
  set(USE_SYSTEM_CPUINFO ON)
@ -487,6 +475,7 @@ if(USE_SYSTEM_LIBS)
  if(USE_NCCL)
    set(USE_SYSTEM_NCCL ON)
  endif()
+  set(USE_SYSTEM_NVTX ON)
 endif()

 # /Z7 override option When generating debug symbols, CMake default to use the
@ -1096,6 +1085,10 @@ if(NOT MSVC)
  append_cxx_flag_if_supported("-fno-math-errno" CMAKE_CXX_FLAGS)
  append_cxx_flag_if_supported("-fno-trapping-math" CMAKE_CXX_FLAGS)
  append_cxx_flag_if_supported("-Werror=format" CMAKE_CXX_FLAGS)
+  if(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 13)
+    append_cxx_flag_if_supported("-Wno-error=dangling-reference" CMAKE_CXX_FLAGS)
+    append_cxx_flag_if_supported("-Wno-error=redundant-move" CMAKE_CXX_FLAGS)
+  endif()
 else()
  # skip unwanted includes from windows.h
  add_compile_definitions(WIN32_LEAN_AND_MEAN)
--- a/8
+++ b/8
@ -116,10 +116,10 @@ aten/src/ATen/detail/MTIAHooksInterface.h @egienvalue
 torch/csrc/mtia/ @egienvalue

 # Profiler
-torch/csrc/autograd/profiler* @aaronenyeshi @sraikund16
-torch/autograd/profiler* @aaronenyeshi @sraikund16
-torch/csrc/profiler/ @aaronenyeshi @sraikund16
-torch/profiler/ @aaronenyeshi @sraikund16
+torch/csrc/autograd/profiler* @sraikund16
+torch/autograd/profiler* @sraikund16
+torch/csrc/profiler/ @sraikund16
+torch/profiler/ @sraikund16

 # AOTDispatch tests
 test/functorch/test_aotdispatch.py @ezyang @Chillee
--- a/aten/src/ATen/Context.h
+++ b/aten/src/ATen/Context.h
@ -39,8 +39,8 @@ class TORCH_API Context {

  const Generator& defaultGenerator(Device device) {
    c10::DeviceType device_type = device.type();
-    initCUDAIfNeeded(device_type);
-    initHIPIfNeeded(device_type);
+    lazyInitDevice(device_type);
+
    if (device_type == at::kCPU) {
      return at::detail::getDefaultCPUGenerator();
    } else if (device_type == at::kCUDA) {
@ -58,6 +58,7 @@ class TORCH_API Context {
      AT_ERROR(c10::DeviceTypeName(device_type), " device type not enabled.");
    }
  }
+
  const AcceleratorHooksInterface& getAcceleratorHooksInterface(
      std::optional<c10::DeviceType> opt_device_type = std::nullopt) {
    c10::DeviceType device_type = opt_device_type.has_value()
@ -80,16 +81,17 @@ class TORCH_API Context {
          c10::DeviceTypeName(device_type), " device type not an accelerator.");
    }
  }
+
  Device getDeviceFromPtr(void* data, c10::DeviceType device_type) {
-    initCUDAIfNeeded(device_type);
-    initHIPIfNeeded(device_type);
-    initXPUIfNeeded(device_type);
+    lazyInitDevice(device_type);
+
    if (device_type == at::kCPU) {
      return c10::DeviceType::CPU;
    } else {
      return getAcceleratorHooksInterface(device_type).getDeviceFromPtr(data);
    }
  }
+
  bool isPinnedPtr(
      const void* data,
      std::optional<c10::DeviceType> device_type = std::nullopt) {
@ -102,10 +104,20 @@ class TORCH_API Context {
    }
    return getAcceleratorHooksInterface(opt_device_type).isPinnedPtr(data);
  }
+
  Allocator* getPinnedMemoryAllocator(
      std::optional<c10::DeviceType> device_type = std::nullopt) {
    return getAcceleratorHooksInterface(device_type).getPinnedMemoryAllocator();
  }
+
+  void lazyInitDevice(c10::DeviceType device_type) {
+    if (device_type != at::kCPU) {
+      c10::call_once(init_[static_cast<int8_t>(device_type)], [&] {
+        getAcceleratorHooksInterface(device_type).init();
+      });
+    }
+  }
+
  static bool hasOpenMP();
  static bool hasMKL();
  static bool hasLAPACK();
@ -158,27 +170,6 @@ class TORCH_API Context {
  static bool hasMAIA() {
    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::MAIA);
  }
-  // defined in header so that getNonVariableType has ability to inline
-  // call_once check. getNonVariableType is called fairly frequently
-  void lazyInitCUDA() {
-    c10::call_once(thc_init, [&] { detail::getCUDAHooks().initCUDA(); });
-  }
-  void lazyInitHIP() {
-    c10::call_once(thh_init, [&] { detail::getHIPHooks().initHIP(); });
-  }
-  void lazyInitXPU() {
-    c10::call_once(thx_init, [&] { detail::getXPUHooks().initXPU(); });
-  }
-  void lazyInitMTIA() {
-    c10::call_once(th_mtia_init, [&] { detail::getMTIAHooks().initMTIA(); });
-  }
-  void lazyInitPrivateUse1() {
-    c10::call_once(thp_init, [&] {
-      if (isPrivateUse1HooksRegistered()) {
-        at::detail::getPrivateUse1Hooks().initPrivateUse1();
-      }
-    });
-  }
  static const at::cuda::NVRTC& getNVRTC() {
    return detail::getCUDAHooks().nvrtc();
  }
@ -353,28 +344,26 @@ class TORCH_API Context {
  bool allowFP16ReductionCPU() const;
  void setAllowFP16ReductionCPU(bool);

+  // Preserved for BC
+  void lazyInitCUDA() {
+    lazyInitDevice(at::kCUDA);
+  }
+  void lazyInitHIP() {
+    lazyInitDevice(at::kHIP);
+  }
+  void lazyInitXPU() {
+    lazyInitDevice(at::kXPU);
+  }
+  void lazyInitMTIA() {
+    lazyInitDevice(at::kMTIA);
+  }
+  void lazyInitPrivateUse1() {
+    lazyInitDevice(at::kPrivateUse1);
+  }
+
 private:
-  void initCUDAIfNeeded(c10::DeviceType p) {
-    if (p == c10::DeviceType::CUDA) {
-      lazyInitCUDA();
-    }
-  }
-  void initHIPIfNeeded(c10::DeviceType p) {
-    if (p == c10::DeviceType::HIP) {
-      lazyInitHIP();
-    }
-  }
-  void initXPUIfNeeded(c10::DeviceType p) {
-    if (p == c10::DeviceType::XPU) {
-      lazyInitXPU();
-    }
-  }
  static bool checkCuBLASConfigDeterministic();
-  c10::once_flag thc_init;
-  c10::once_flag thh_init;
-  c10::once_flag thx_init;
-  c10::once_flag th_mtia_init;
-  c10::once_flag thp_init;
+  std::array<c10::once_flag, at::COMPILE_TIME_MAX_DEVICE_TYPES> init_;
  bool enabled_cudnn = true;
  bool deterministic_cudnn = false;
  bool deterministic_mkldnn = false;
--- a/aten/src/ATen/DLConvertor.cpp
+++ b/aten/src/ATen/DLConvertor.cpp
@ -22,6 +22,13 @@ DLDataType getDLDataType(const Tensor& t) {
    case ScalarType::UInt64:
      dtype.code = DLDataTypeCode::kDLUInt;
      break;
+    case ScalarType::Int1:
+    case ScalarType::Int2:
+    case ScalarType::Int3:
+    case ScalarType::Int4:
+    case ScalarType::Int5:
+    case ScalarType::Int6:
+    case ScalarType::Int7:
    case ScalarType::Char:
      dtype.code = DLDataTypeCode::kDLInt;
      break;
@ -49,11 +56,7 @@ DLDataType getDLDataType(const Tensor& t) {
      dtype.code = DLDataTypeCode::kDLBool;
      break;
    case ScalarType::ComplexHalf:
-      dtype.code = DLDataTypeCode::kDLComplex;
-      break;
    case ScalarType::ComplexFloat:
-      dtype.code = DLDataTypeCode::kDLComplex;
-      break;
    case ScalarType::ComplexDouble:
      dtype.code = DLDataTypeCode::kDLComplex;
      break;
@ -90,7 +93,7 @@ DLDataType getDLDataType(const Tensor& t) {

 static DLDevice getDLDevice(const Tensor& tensor, c10::DeviceIndex device_id) {
  DLDevice ctx;
-  ctx.device_id = static_cast<int32_t>(device_id);
+  ctx.device_id = static_cast<int32_t>(static_cast<unsigned char>(device_id));
  switch (tensor.device().type()) {
    case DeviceType::CPU:
      ctx.device_type = DLDeviceType::kDLCPU;
@ -253,10 +256,12 @@ ScalarType toScalarType(const DLDataType& dtype) {
 }

 // NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
+namespace {
 struct ATenDLMTensor {
  Tensor handle;
-  DLManagedTensor tensor;
+  DLManagedTensor tensor{};
 };
+} // namespace

 static void deleter(DLManagedTensor* arg) {
  delete static_cast<ATenDLMTensor*>(arg->manager_ctx);
--- a/aten/src/ATen/Dispatch.h
+++ b/aten/src/ATen/Dispatch.h
@ -78,8 +78,8 @@ TORCH_API void record_kernel_function_dtype(std::string name);
    AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type);                      \
    using scalar_t = scalar_type;                                     \
    using underlying_t C10_UNUSED = typename scalar_t::underlying;    \
-    const auto& SCALAR_TYPE C10_UNUSED = enum_type;                   \
-    const auto& UNDERLYING_TYPE C10_UNUSED = toUnderlying(enum_type); \
+    C10_UNUSED const auto& SCALAR_TYPE = enum_type;                   \
+    C10_UNUSED const auto& UNDERLYING_TYPE = toUnderlying(enum_type); \
    return __VA_ARGS__();                                             \
  }

@ -89,8 +89,8 @@ TORCH_API void record_kernel_function_dtype(std::string name);
    AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type);                      \
    using scalar_t = scalar_type;                                     \
    using underlying_t C10_UNUSED = typename scalar_t::underlying;    \
-    const auto& SCALAR_TYPE C10_UNUSED = enum_type;                   \
-    const auto& UNDERLYING_TYPE C10_UNUSED = toUnderlying(enum_type); \
+    C10_UNUSED const auto& SCALAR_TYPE = enum_type;                   \
+    C10_UNUSED const auto& UNDERLYING_TYPE = toUnderlying(enum_type); \
    C10_UNUSED int bit_width = bitwidth;                              \
    C10_UNUSED int64_t quant_min = qmin;                              \
    C10_UNUSED int64_t quant_max = qmax;                              \
--- a/aten/src/ATen/Dispatch_v2.h
+++ b/aten/src/ATen/Dispatch_v2.h
@ -112,12 +112,12 @@

 // Ensure we never have too many scalar types for the expansion here to
 // support.  To bump this, you must regenerate the macros below.
-static_assert(static_cast<int>(c10::ScalarType::NumOptions) < 45);
+static_assert(static_cast<int>(c10::ScalarType::NumOptions) < 60);

 // Python code to regenerate generate code below:
 #if 0

-num_args = 45
+num_args = 60

 nums = ', '.join(str(i) for i in reversed(range(num_args+1)))
 args = ', '.join(f'_{i}' for i in range(1, num_args+1))
@ -135,8 +135,8 @@ for i in range(1, num_args+1):
 // Begin generated code
 // clang-format off

-#define AT_NUM_ARGS(...) AT_EXPAND(AT_NUM_ARGS_AUX(__VA_ARGS__, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0))
-#define AT_NUM_ARGS_AUX(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, N, ...) N
+#define AT_NUM_ARGS(...) AT_EXPAND(AT_NUM_ARGS_AUX(__VA_ARGS__, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0))
+#define AT_NUM_ARGS_AUX(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, N, ...) N
 #define AT_AP1(N, _1) AT_DISPATCH_CASE(_1, N)
 #define AT_AP2(N, _1, _2) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N)
 #define AT_AP3(N, _1, _2, _3) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N)
@ -182,5 +182,21 @@ for i in range(1, num_args+1):
 #define AT_AP43(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N)
 #define AT_AP44(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N)
 #define AT_AP45(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N)
+#define AT_AP46(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N)
+#define AT_AP47(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N)
+#define AT_AP48(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N)
+#define AT_AP49(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N)
+#define AT_AP50(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N)
+#define AT_AP51(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N)
+#define AT_AP52(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N)
+#define AT_AP53(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N)
+#define AT_AP54(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N)
+#define AT_AP55(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N) AT_DISPATCH_CASE(_55, N)
+#define AT_AP56(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N) AT_DISPATCH_CASE(_55, N) AT_DISPATCH_CASE(_56, N)
+#define AT_AP57(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N) AT_DISPATCH_CASE(_55, N) AT_DISPATCH_CASE(_56, N) AT_DISPATCH_CASE(_57, N)
+#define AT_AP58(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N) AT_DISPATCH_CASE(_55, N) AT_DISPATCH_CASE(_56, N) AT_DISPATCH_CASE(_57, N) AT_DISPATCH_CASE(_58, N)
+#define AT_AP59(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N) AT_DISPATCH_CASE(_55, N) AT_DISPATCH_CASE(_56, N) AT_DISPATCH_CASE(_57, N) AT_DISPATCH_CASE(_58, N) AT_DISPATCH_CASE(_59, N)
+#define AT_AP60(N, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N) AT_DISPATCH_CASE(_4, N) AT_DISPATCH_CASE(_5, N) AT_DISPATCH_CASE(_6, N) AT_DISPATCH_CASE(_7, N) AT_DISPATCH_CASE(_8, N) AT_DISPATCH_CASE(_9, N) AT_DISPATCH_CASE(_10, N) AT_DISPATCH_CASE(_11, N) AT_DISPATCH_CASE(_12, N) AT_DISPATCH_CASE(_13, N) AT_DISPATCH_CASE(_14, N) AT_DISPATCH_CASE(_15, N) AT_DISPATCH_CASE(_16, N) AT_DISPATCH_CASE(_17, N) AT_DISPATCH_CASE(_18, N) AT_DISPATCH_CASE(_19, N) AT_DISPATCH_CASE(_20, N) AT_DISPATCH_CASE(_21, N) AT_DISPATCH_CASE(_22, N) AT_DISPATCH_CASE(_23, N) AT_DISPATCH_CASE(_24, N) AT_DISPATCH_CASE(_25, N) AT_DISPATCH_CASE(_26, N) AT_DISPATCH_CASE(_27, N) AT_DISPATCH_CASE(_28, N) AT_DISPATCH_CASE(_29, N) AT_DISPATCH_CASE(_30, N) AT_DISPATCH_CASE(_31, N) AT_DISPATCH_CASE(_32, N) AT_DISPATCH_CASE(_33, N) AT_DISPATCH_CASE(_34, N) AT_DISPATCH_CASE(_35, N) AT_DISPATCH_CASE(_36, N) AT_DISPATCH_CASE(_37, N) AT_DISPATCH_CASE(_38, N) AT_DISPATCH_CASE(_39, N) AT_DISPATCH_CASE(_40, N) AT_DISPATCH_CASE(_41, N) AT_DISPATCH_CASE(_42, N) AT_DISPATCH_CASE(_43, N) AT_DISPATCH_CASE(_44, N) AT_DISPATCH_CASE(_45, N) AT_DISPATCH_CASE(_46, N) AT_DISPATCH_CASE(_47, N) AT_DISPATCH_CASE(_48, N) AT_DISPATCH_CASE(_49, N) AT_DISPATCH_CASE(_50, N) AT_DISPATCH_CASE(_51, N) AT_DISPATCH_CASE(_52, N) AT_DISPATCH_CASE(_53, N) AT_DISPATCH_CASE(_54, N) AT_DISPATCH_CASE(_55, N) AT_DISPATCH_CASE(_56, N) AT_DISPATCH_CASE(_57, N) AT_DISPATCH_CASE(_58, N) AT_DISPATCH_CASE(_59, N) AT_DISPATCH_CASE(_60, N)
+
 // End generated code
 // clang-format on
--- a/aten/src/ATen/ZeroTensorFallback.cpp
+++ b/aten/src/ATen/ZeroTensorFallback.cpp
@ -66,8 +66,7 @@ namespace at {
      } else if (ivalue.isTensorList()) {
        auto tensors = std::move(ivalue).toTensorList();
        for(const auto j : c10::irange(tensors.size())) {
-          const auto& tensor_ref = tensors[j];
-          const Tensor& tensor = tensor_ref;
+          const Tensor& tensor = tensors[j];
          if (tensor._is_zerotensor()) {
            // TODO: assert requires_grad=False
            //_like should not propagate zerotensor dispatch key
--- a/aten/src/ATen/core/Dict.h
+++ b/aten/src/ATen/core/Dict.h
@ -314,7 +314,7 @@ public:
   *
   * @return The number of elements removed. This is either '1' if an element with the key existed, or '0' if it didn't.
   */
-  C10_NODISCARD size_t erase(const Key& key) const;
+  [[nodiscard]] size_t erase(const Key& key) const;

  /**
   * Returns the mapped value of the element with key equivalent to key.
--- a/aten/src/ATen/core/Dict_inl.h
+++ b/aten/src/ATen/core/Dict_inl.h
@ -142,8 +142,8 @@ void Dict<Key, Value>::erase(iterator iter) const {
  impl_->dict.erase(iter.entryRef_.iterator_);
 }

-template<class Key, class Value>
-C10_NODISCARD size_t Dict<Key, Value>::erase(const Key& key) const {
+template <class Key, class Value>
+[[nodiscard]] size_t Dict<Key, Value>::erase(const Key& key) const {
  return impl_->dict.erase(key);
 }

--- a/aten/src/ATen/core/IListRef_inl.h
+++ b/aten/src/ATen/core/IListRef_inl.h
@ -168,8 +168,7 @@ class IListRefTagImpl<IListRefTag::Boxed, at::OptionalTensorRef>
   */
  static IListRefConstRef<at::OptionalTensorRef> iterator_get(
      const typename list_type::const_iterator& it) {
-    const auto& elem = *it;
-    const auto& ivalue = elem.get();
+    const auto& ivalue = (*it).get();
    if (!ivalue.isNone()) {
        const auto& tensor = ivalue.toTensor();
        return (tensor.defined()) ? tensor : at::OptionalTensorRef{};
--- a/aten/src/ATen/core/dispatch/DispatchKeyExtractor.h
+++ b/aten/src/ATen/core/dispatch/DispatchKeyExtractor.h
@ -151,9 +151,7 @@ public:
        // no safe toTensorRef method, alas)
        ks = ks | ivalue.unsafeToTensorImpl()->key_set();
      } else if (C10_UNLIKELY(ivalue.isTensorList())) {
-        const c10::List<at::Tensor> tensorlist = ivalue.toTensorList();
-        for (const auto& tensor_ref : tensorlist) {
-          const at::Tensor& tensor = tensor_ref;
+        for (const at::Tensor& tensor : ivalue.toTensorList()) {
          ks = ks | tensor.key_set();
        }
      }
--- a/aten/src/ATen/core/function_schema.h
+++ b/aten/src/ATen/core/function_schema.h
@ -108,7 +108,7 @@ struct TORCH_API Argument {
    return is_out_;
  }

-  C10_NODISCARD const AliasInfo* alias_info() const {
+  [[nodiscard]] const AliasInfo* alias_info() const {
    return alias_info_.get();
  }

--- a/aten/src/ATen/core/ivalue.h
+++ b/aten/src/ATen/core/ivalue.h
@ -522,7 +522,7 @@ struct TORCH_API IValue final {
  }
  c10::intrusive_ptr<ivalue::Tuple> toTuple() &&;
  c10::intrusive_ptr<ivalue::Tuple> toTuple() const&;
-  C10_NODISCARD ivalue::Tuple& toTupleRef() const;
+  [[nodiscard]] ivalue::Tuple& toTupleRef() const;

  // Double
  IValue(double d) : tag(Tag::Double) {
--- a/aten/src/ATen/core/ivalue_inl.h
+++ b/aten/src/ATen/core/ivalue_inl.h
@ -500,7 +500,7 @@ struct TORCH_API TupleElements {
    return *this;
  }

-  C10_NODISCARD c10::ArrayRef<IValue> asArrayRef() const {
+  [[nodiscard]] c10::ArrayRef<IValue> asArrayRef() const {
    if (inlineSize_) {
      return c10::ArrayRef<IValue>(elementsInline_, inlineSize_);
    } else {
@ -527,15 +527,15 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD bool empty() const {
+  [[nodiscard]] bool empty() const {
    return inlineSize_ ? false : elementsVector_.empty();
  }

-  C10_NODISCARD size_t size() const {
+  [[nodiscard]] size_t size() const {
    return inlineSize_ ? inlineSize_ : elementsVector_.size();
  }

-  C10_NODISCARD IValue& operator[](size_t idx) {
+  [[nodiscard]] IValue& operator[](size_t idx) {
    if (inlineSize_) {
      return elementsInline_[idx];
    } else {
@ -543,7 +543,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD const IValue& operator[](size_t idx) const {
+  [[nodiscard]] const IValue& operator[](size_t idx) const {
    if (inlineSize_) {
      return elementsInline_[idx];
    } else {
@ -551,7 +551,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD IValue& at(size_t idx) {
+  [[nodiscard]] IValue& at(size_t idx) {
    if (inlineSize_) {
      TORCH_INTERNAL_ASSERT_DEBUG_ONLY(inlineSize_ <= 3);
      TORCH_CHECK(idx < inlineSize_, "TupleElements: invalid index Index = ", idx, "; Length = ", inlineSize_);
@ -561,7 +561,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD const IValue& at(size_t idx) const {
+  [[nodiscard]] const IValue& at(size_t idx) const {
    if (inlineSize_) {
      TORCH_INTERNAL_ASSERT_DEBUG_ONLY(inlineSize_ <= 3);
      TORCH_CHECK(idx < inlineSize_, "TupleElements: invalid index Index = ", idx, "; Length = ", inlineSize_);
@ -572,7 +572,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD iterator begin() {
+  [[nodiscard]] iterator begin() {
    if (inlineSize_) {
      return elementsInline_;
    } else {
@ -580,7 +580,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD iterator end() {
+  [[nodiscard]] iterator end() {
    if (inlineSize_) {
      return elementsInline_ + inlineSize_;
    } else {
@ -588,7 +588,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD const_iterator begin() const {
+  [[nodiscard]] const_iterator begin() const {
    if (inlineSize_) {
      return elementsInline_;
    } else {
@ -596,7 +596,7 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD const_iterator end() const {
+  [[nodiscard]] const_iterator end() const {
    if (inlineSize_) {
      return elementsInline_ + inlineSize_;
    } else {
@ -604,27 +604,27 @@ struct TORCH_API TupleElements {
    }
  }

-  C10_NODISCARD const_iterator cbegin() const {
+  [[nodiscard]] const_iterator cbegin() const {
    return begin();
  }

-  C10_NODISCARD const_iterator cend() const {
+  [[nodiscard]] const_iterator cend() const {
    return end();
  }

-  C10_NODISCARD std::vector<IValue> vec() const & {
+  [[nodiscard]] std::vector<IValue> vec() const& {
    return asArrayRef().vec();
  }

-  C10_NODISCARD IValue& back() {
+  [[nodiscard]] IValue& back() {
    return *(end() - 1);
  }

-  C10_NODISCARD const IValue& back() const {
+  [[nodiscard]] const IValue& back() const {
    return *(end() - 1);
  }

-  C10_NODISCARD std::vector<IValue> vec() && {
+  [[nodiscard]] std::vector<IValue> vec() && {
    std::vector<IValue> result;
    result.reserve(size());
    for (auto&& iv : *this) {
--- a/aten/src/ATen/cpu/vec/vec256/vec256_convert.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_convert.h
@ -271,9 +271,9 @@ struct VecConvert<
    1,
    int64_t,
    2,
-    typename std::enable_if<
+    std::enable_if_t<
        std::is_same_v<dst_t, int8_t> ||
-        std::is_same_v<dst_t, uint8_t>>::type> {
+        std::is_same_v<dst_t, uint8_t>>> {
  static inline VectorizedN<dst_t, 1> apply(
      const VectorizedN<int64_t, 2>& src) {
    return VecConvert<dst_t, 1, int32_t, 1>::apply(
--- a/aten/src/ATen/cpu/vec/vec256/zarch/vec256_zarch.h
+++ b/aten/src/ATen/cpu/vec/vec256/zarch/vec256_zarch.h
@ -1190,8 +1190,8 @@ struct Vectorized<T, std::enable_if_t<is_zarch_implemented<T>()>> {
      typename U = T,
      std::enable_if_t<std::is_same<U, double>::value, int> = 0>
  Vectorized<T> swapped() const {
-    vtype v0 = vec_permi(_vec0, _vec0, 2);
-    vtype v1 = vec_permi(_vec1, _vec1, 2);
+    vtype v0 = {_vec0[1], _vec0[0]};
+    vtype v1 = {_vec1[1], _vec1[0]};
    return {v0, v1};
  }

@ -1685,6 +1685,7 @@ std::pair<Vectorized<V>, Vectorized<V>> unpack(const Vectorized<T>& x) {
  return {Vectorized<V>{vec0, vec1}, Vectorized<V>{vec2, vec3}};
 }

+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-function")
 template <>
 std::pair<Vectorized<int16_t>, Vectorized<int16_t>> unpack<uint8_t, int16_t>(
    const Vectorized<uint8_t>& x) {
@ -1702,6 +1703,7 @@ std::pair<Vectorized<int16_t>, Vectorized<int16_t>> unpack<uint8_t, int16_t>(
      cast_zvector<uint16_t, int16_t>(Vectorized<uint16_t>{vec0, vec1}),
      cast_zvector<uint16_t, int16_t>(Vectorized<uint16_t>{vec2, vec3})};
 }
+C10_DIAGNOSTIC_POP()

 template <typename T, typename V = typename pack_type<T>::type>
 Vectorized<V> pack(const Vectorized<T>& first, const Vectorized<T>& second) {
@ -1710,6 +1712,7 @@ Vectorized<V> pack(const Vectorized<T>& first, const Vectorized<T>& second) {
  return Vectorized<V>{vec0, vec1};
 }

+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-function")
 template <>
 Vectorized<uint8_t> pack(
    const Vectorized<int16_t>& first,
@ -1718,6 +1721,7 @@ Vectorized<uint8_t> pack(
  auto vec1 = vec_packsu(second.vec0(), second.vec1());
  return Vectorized<uint8_t>{vec0, vec1};
 }
+C10_DIAGNOSTIC_POP()

 } /* unnamed namespace */

@ -1735,7 +1739,7 @@ struct Vectorized<T, std::enable_if_t<is_zarch_implemented_quant<T>()>> {
    return VECTOR_WIDTH / sizeof(value_type);
  }

-  static constexpr size_t float_num_vecs() {
+  static constexpr int float_num_vecs() {
    return size() / Vectorized<float>::size();
  }
  static constexpr int int_num_vecs() {
@ -2419,8 +2423,8 @@ struct Vectorized<T, std::enable_if_t<is_zarch_implemented_complex<T>()>> {
  static typename Vectorized<T>::vinner_type real_neg(const typename Vectorized<T>::vinner_type &a)
  {
    auto a_neg = a.neg();
-    auto v0 = vec_permi(a_neg.vec0(), a.vec0(), 1);
-    auto v1 = vec_permi(a_neg.vec1(), a.vec1(), 1);
+    vtype v0 = {a_neg.vec0()[0], a.vec0()[1]};
+    vtype v1 = {a_neg.vec1()[0], a.vec1()[1]};
    return { v0, v1 };
  }

@ -2732,10 +2736,10 @@ std::pair<Vectorized<T>, Vectorized<T>> inline inner_interleave2(
  //   a      = {a0, a1, a2, a3}
  //   b      = {b0, b1, b2, b3}
  using vtype = typename Vectorized<T>::vtype;
-  vtype ab00 = vec_permi(a.vec0(), b.vec0(), 0);
-  vtype ab11 = vec_permi(a.vec0(), b.vec0(), 3);
-  vtype ab2_00 = vec_permi(a.vec1(), b.vec1(), 0);
-  vtype ab2_11 = vec_permi(a.vec1(), b.vec1(), 3);
+  vtype ab00 = {a.vec0()[0], b.vec0()[0]};
+  vtype ab11 = {a.vec0()[1], b.vec0()[1]};
+  vtype ab2_00 = {a.vec1()[0], b.vec1()[0]};
+  vtype ab2_11 = {a.vec1()[1], b.vec1()[1]};
  //   return {a0, b0, a1, b1}
  //          {a2, b2, a3, b3}
  return std::make_pair(
@ -2750,11 +2754,11 @@ std::pair<Vectorized<T>, Vectorized<T>> inline inner_deinterleave2(
  //   a = {a0, b0, a1, b1}
  //   b = {a2, b2, a3, b3}
  using vtype = typename Vectorized<T>::vtype;
-  vtype aa01 = vec_permi(a.vec0(), a.vec1(), 0);
-  vtype aa23 = vec_permi(b.vec0(), b.vec1(), 0);
+  vtype aa01 = {a.vec0()[0], a.vec1()[0]};
+  vtype aa23 = {b.vec0()[0], b.vec1()[0]};

-  vtype bb_01 = vec_permi(a.vec0(), a.vec1(), 3);
-  vtype bb_23 = vec_permi(b.vec0(), b.vec1(), 3);
+  vtype bb_01 = {a.vec0()[1], a.vec1()[1]};
+  vtype bb_23 = {b.vec0()[1], b.vec1()[1]};

  // swap lanes:
  //   return {a0, a1, a2, a3}
@ -2868,7 +2872,7 @@ std::pair<Vectorized<int64_t>, Vectorized<int64_t>> inline deinterleave2<
 }

 template <typename T>
-typename std::enable_if<std::is_same<T, uint8_t>::value, at::vec::Vectorized<float>>::type
+std::enable_if_t<std::is_same_v<T, uint8_t>, at::vec::Vectorized<float>>
 inline convert_int8_to_float(const Vectorized<T> &src) {
  // Note: this function only convert inputs number of elements equal to at::vec::Vectorized<float>.size()
  // Only handle first 64 bits
@ -2878,7 +2882,7 @@ inline convert_int8_to_float(const Vectorized<T> &src) {
 }

 template <typename T>
-typename std::enable_if<std::is_same<T, uint8_t>::value, at::vec::Vectorized<T>>::type
+std::enable_if_t<std::is_same_v<T, uint8_t>, at::vec::Vectorized<T>>
 inline convert_float_to_int8(const Vectorized<float> &src) {
  constexpr auto min_val = std::numeric_limits<T>::min();
  constexpr auto max_val = std::numeric_limits<T>::max();
--- a/aten/src/ATen/cpu/vec/vec512/vec512_convert.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512_convert.h
@ -281,9 +281,9 @@ struct VecConvert<
    1,
    int64_t,
    2,
-    typename std::enable_if<
+    std::enable_if_t<
        std::is_same_v<dst_t, int8_t> ||
-        std::is_same_v<dst_t, uint8_t>>::type> {
+        std::is_same_v<dst_t, uint8_t>>> {
  static inline VectorizedN<dst_t, 1> apply(
      const VectorizedN<int64_t, 2>& src) {
    return VecConvert<dst_t, 1, int32_t, 1>::apply(
--- a/aten/src/ATen/cpu/vec/vec512/vec512_mask.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512_mask.h
@ -84,9 +84,9 @@ struct VecMaskLoad<
    dst_n,
    mask_t,
    dst_n,
-    typename std::enable_if<
+    std::enable_if_t<
        std::is_same_v<data_t, BFloat16> ||
-        std::is_same_v<data_t, Half>>::type> {
+        std::is_same_v<data_t, Half>>> {
  static inline VectorizedN<data_t, dst_n> apply(
      const data_t* ptr,
      const VecMask<mask_t, dst_n>& vec_mask) {
@ -151,9 +151,9 @@ struct VecMaskLoad<
    1,
    mask_t,
    1,
-    typename std::enable_if<
+    std::enable_if_t<
        std::is_same_v<data_t, int8_t> ||
-        std::is_same_v<data_t, uint8_t>>::type> {
+        std::is_same_v<data_t, uint8_t>>> {
  static inline VectorizedN<data_t, 1> apply(
      const data_t* ptr,
      const VecMask<mask_t, 1>& vec_mask) {
@ -173,9 +173,9 @@ struct VecMaskLoad<
    2,
    mask_t,
    1,
-    typename std::enable_if<
+    std::enable_if_t<
        std::is_same_v<data_t, int64_t> ||
-        std::is_same_v<data_t, double>>::type> {
+        std::is_same_v<data_t, double>>> {
  static inline VectorizedN<data_t, 2> apply(
      const data_t* ptr,
      const VecMask<mask_t, 1>& vec_mask) {
--- a/aten/src/ATen/cuda/EmptyTensor.cpp
+++ b/aten/src/ATen/cuda/EmptyTensor.cpp
@ -10,7 +10,7 @@ TensorBase empty_cuda(
    ScalarType dtype,
    std::optional<Device> device_opt,
    std::optional<c10::MemoryFormat> memory_format_opt) {
-  at::globalContext().lazyInitCUDA();
+  at::globalContext().lazyInitDevice(c10::DeviceType::CUDA);
  const auto device = device_or_default(device_opt);
  TORCH_INTERNAL_ASSERT(device.is_cuda());
  const DeviceGuard device_guard(device);
@ -50,7 +50,7 @@ TensorBase empty_strided_cuda(
    IntArrayRef stride,
    ScalarType dtype,
    std::optional<Device> device_opt) {
-  at::globalContext().lazyInitCUDA();
+  at::globalContext().lazyInitDevice(c10::DeviceType::CUDA);
  const auto device = device_or_default(device_opt);
  TORCH_INTERNAL_ASSERT(device.is_cuda());
  const DeviceGuard device_guard(device);
--- a/aten/src/ATen/cuda/Exceptions.h
+++ b/aten/src/ATen/cuda/Exceptions.h
@ -162,7 +162,7 @@ constexpr const char* _cusolver_backend_suggestion =            \
    CUresult __err = EXPR;                                                                                       \
    if (__err != CUDA_SUCCESS) {                                                                                 \
      const char* err_str;                                                                                       \
-      CUresult get_error_str_err C10_UNUSED = at::globalContext().getNVRTC().cuGetErrorString(__err, &err_str);  \
+      C10_UNUSED CUresult get_error_str_err = at::globalContext().getNVRTC().cuGetErrorString(__err, &err_str);  \
      if (get_error_str_err != CUDA_SUCCESS) {                                                                   \
        AT_ERROR("CUDA driver error: unknown error");                                                            \
      } else {                                                                                                   \
--- a/aten/src/ATen/cuda/PeerToPeerAccess.cpp
+++ b/aten/src/ATen/cuda/PeerToPeerAccess.cpp
@ -34,7 +34,7 @@ void init_p2p_access_cache(int64_t num_devices) {
 }  // namespace detail

 bool get_p2p_access(int dev, int dev_to_access) {
-  at::globalContext().lazyInitCUDA();
+  at::globalContext().lazyInitDevice(c10::DeviceType::CUDA);

  TORCH_CHECK(dev >= 0 || dev < num_devices_,
              dev, " is not a device");
--- a/aten/src/ATen/cuda/detail/CUDAHooks.cpp
+++ b/aten/src/ATen/cuda/detail/CUDAHooks.cpp
@ -84,7 +84,7 @@ struct _Initializer {
 // NB: deleter is dynamic, because we need it to live in a separate
 // compilation unit (alt is to have another method in hooks, but
 // let's not if we don't need to!)
-void CUDAHooks::initCUDA() const {
+void CUDAHooks::init() const {
  C10_LOG_API_USAGE_ONCE("aten.init.cuda");
  // Force the update to enable unit testing. This code get executed before unit tests
  // have a chance to enable vitals.
--- a/aten/src/ATen/cuda/detail/CUDAHooks.h
+++ b/aten/src/ATen/cuda/detail/CUDAHooks.h
@ -19,7 +19,7 @@ TORCH_CUDA_CPP_API void set_magma_init_fn(void (*magma_init_fn)());
 // The real implementation of CUDAHooksInterface
 struct CUDAHooks : public at::CUDAHooksInterface {
  CUDAHooks(at::CUDAHooksArgs) {}
-  void initCUDA() const override;
+  void init() const override;
  Device getDeviceFromPtr(void* data) const override;
  bool isPinnedPtr(const void* data) const override;
  const Generator& getDefaultCUDAGenerator(DeviceIndex device_index = -1) const override;
--- a/aten/src/ATen/detail/AcceleratorHooksInterface.h
+++ b/aten/src/ATen/detail/AcceleratorHooksInterface.h
@ -19,6 +19,10 @@ struct TORCH_API AcceleratorHooksInterface {
  // Whether the device at device_index is fully initialized or not.
  virtual bool hasPrimaryContext(DeviceIndex device_index) const = 0;

+  virtual void init() const {
+    TORCH_CHECK(false, "Backend doesn`t support init()");
+  }
+
  virtual DeviceIndex deviceCount() const {
    return 0;
  }
--- a/aten/src/ATen/detail/CUDAHooksInterface.h
+++ b/aten/src/ATen/detail/CUDAHooksInterface.h
@ -65,7 +65,7 @@ struct TORCH_API CUDAHooksInterface : AcceleratorHooksInterface {
  ~CUDAHooksInterface() override = default;

  // Initialize THCState and, transitively, the CUDA state
-  virtual void initCUDA() const {
+  void init() const override {
    TORCH_CHECK(false, "Cannot initialize CUDA without ATen_cuda library. ", CUDA_HELP);
  }

--- a/aten/src/ATen/detail/HIPHooksInterface.h
+++ b/aten/src/ATen/detail/HIPHooksInterface.h
@ -26,9 +26,8 @@ struct TORCH_API HIPHooksInterface : AcceleratorHooksInterface {
  // squelch -Werror=non-virtual-dtor
  ~HIPHooksInterface() override = default;

-  // Initialize the HIP library state
-  virtual void initHIP() const {
-    AT_ERROR("Cannot initialize HIP without ATen_hip library.");
+  void init() const override {
+    TORCH_CHECK(false, "Cannot initialize HIP without ATen_hip library.");
  }

  virtual std::unique_ptr<c10::GeneratorImpl> initHIPGenerator(Context*) const {
--- a/aten/src/ATen/detail/IPUHooksInterface.h
+++ b/aten/src/ATen/detail/IPUHooksInterface.h
@ -1,14 +1,25 @@
 #pragma once

 #include <ATen/core/Generator.h>
+#include <ATen/detail/AcceleratorHooksInterface.h>
+
 #include <c10/core/Allocator.h>
 #include <c10/util/Exception.h>
 #include <c10/util/Registry.h>

 namespace at {

-struct TORCH_API IPUHooksInterface {
-  virtual ~IPUHooksInterface() = default;
+struct TORCH_API IPUHooksInterface: AcceleratorHooksInterface {
+  ~IPUHooksInterface() override = default;
+
+  void init() const override {
+    TORCH_CHECK(false, "Cannot initialize IPU without ATen_ipu library.");
+  }
+
+  bool hasPrimaryContext(DeviceIndex device_index) const override {
+    TORCH_CHECK(false, "Cannot initialize IPU without ATen_ipu library.");
+    return false;
+  }

  virtual const Generator& getDefaultIPUGenerator(
      DeviceIndex device_index [[maybe_unused]] = -1) const {
--- a/aten/src/ATen/detail/MAIAHooksInterface.h
+++ b/aten/src/ATen/detail/MAIAHooksInterface.h
@ -3,13 +3,24 @@
 #include <c10/util/Exception.h>
 #include <c10/util/Registry.h>

+#include <ATen/detail/AcceleratorHooksInterface.h>
+
 // NB: Class must live in `at` due to limitations of Registry.h.
 namespace at {

-struct TORCH_API MAIAHooksInterface {
+struct TORCH_API MAIAHooksInterface : AcceleratorHooksInterface {
  // This should never actually be implemented, but it is used to
  // squelch -Werror=non-virtual-dtor
-  virtual ~MAIAHooksInterface() = default;
+  ~MAIAHooksInterface() override = default;
+
+  void init() const override {
+    TORCH_CHECK(false, "Cannot initialize MAIA without ATen_maia library.");
+  }
+
+  bool hasPrimaryContext(DeviceIndex device_index) const override {
+    TORCH_CHECK(false, "Cannot initialize MAIA without ATen_maia library.");
+    return false;
+  }

  virtual std::string showConfig() const {
    TORCH_CHECK(false, "Cannot query detailed MAIA version information.");
--- a/aten/src/ATen/detail/MPSHooksInterface.h
+++ b/aten/src/ATen/detail/MPSHooksInterface.h
@ -22,7 +22,7 @@ struct TORCH_API MPSHooksInterface : AcceleratorHooksInterface {
  ~MPSHooksInterface() override = default;

  // Initialize the MPS library state
-  virtual void initMPS() const {
+  void init() const override {
    FAIL_MPSHOOKS_FUNC(__func__);
  }
  virtual bool hasMPS() const {
--- a/aten/src/ATen/detail/MTIAHooksInterface.h
+++ b/aten/src/ATen/detail/MTIAHooksInterface.h
@ -31,7 +31,7 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {

  ~MTIAHooksInterface() override = default;

-  virtual void initMTIA() const {
+  void init() const override {
    // Avoid logging here, since MTIA needs init devices first then it will know
    // how many devices are available. Make it as no-op if mtia extension is not
    // dynamically loaded.
--- a/aten/src/ATen/detail/PrivateUse1HooksInterface.h
+++ b/aten/src/ATen/detail/PrivateUse1HooksInterface.h
@ -40,7 +40,7 @@ struct TORCH_API PrivateUse1HooksInterface : AcceleratorHooksInterface {
        "You should register `PrivateUse1HooksInterface` for PrivateUse1 before call `hasPrimaryContext`.");
  }

-  virtual void initPrivateUse1() const {}
+  void init() const override {}
  virtual void resizePrivateUse1Bytes(
      const c10::Storage& storage,
      size_t newsize) const {
--- a/aten/src/ATen/detail/XPUHooksInterface.h
+++ b/aten/src/ATen/detail/XPUHooksInterface.h
@ -14,10 +14,8 @@ namespace at {
 struct TORCH_API XPUHooksInterface : AcceleratorHooksInterface{
  ~XPUHooksInterface() override = default;

-  virtual void initXPU() const {
-    TORCH_CHECK(
-        false,
-        "Cannot initialize XPU without ATen_xpu library.");
+  void init() const override {
+    TORCH_CHECK(false, "Cannot initialize XPU without ATen_xpu library.");
  }

  virtual bool hasXPU() const {
--- a/aten/src/ATen/mps/MPSDevice.mm
+++ b/aten/src/ATen/mps/MPSDevice.mm
@ -36,7 +36,9 @@ id<MTLLibrary> MPSDevice::getMetalIndexingLibrary() {
    if (isMacOS13Plus(MacOSVersion::MACOS_VER_15_0_PLUS)) {
      options.mathMode = MTLMathModeFast;
    } else {
+      C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wdeprecated-declarations")
      [options setFastMathEnabled:YES];
+      C10_DIAGNOSTIC_POP()
    }
    _mtl_indexing_library = [_mtl_device newLibraryWithSource:[NSString stringWithCString:mps::indexing_metal_shaders
                                                                                 encoding:NSASCIIStringEncoding]
--- a/aten/src/ATen/mps/MPSHooks.h
+++ b/aten/src/ATen/mps/MPSHooks.h
@ -12,7 +12,7 @@ namespace at::mps {
 // The real implementation of MPSHooksInterface
 struct MPSHooks : public at::MPSHooksInterface {
  MPSHooks(at::MPSHooksArgs) {}
-  void initMPS() const override;
+  void init() const override;

  // MPSDevice interface
  bool hasMPS() const override;
--- a/aten/src/ATen/mps/MPSHooks.mm
+++ b/aten/src/ATen/mps/MPSHooks.mm
@ -10,7 +10,7 @@

 namespace at::mps {

-void MPSHooks::initMPS() const {
+void MPSHooks::init() const {
  C10_LOG_API_USAGE_ONCE("aten.init.mps");
  // TODO: initialize MPS devices and streams here
 }
--- a/aten/src/ATen/native/EmbeddingBag.cpp
+++ b/aten/src/ATen/native/EmbeddingBag.cpp
@ -106,7 +106,7 @@ bool is_fast_path(const Tensor& src, const std::optional<Tensor>& scale, Tensor&
 // index_add (using add_indices as the index), without creating an intermediary
 // tensor to hold the selected embeddings
 template <typename data_t, typename index_t>
-static typename std::enable_if<std::is_same<data_t, double>::value, void>::type
+static std::enable_if_t<std::is_same_v<data_t, double>, void>
 index_select_add(
    const Tensor& select_indices,
    const Tensor& add_indices,
@ -184,10 +184,9 @@ void fbgemm_spmdm_report_error_(
 } // namespace

 template <typename data_t, typename index_t>
-typename std::enable_if<
-    std::is_same<data_t, at::Half>::value ||
-        std::is_same<data_t, at::BFloat16>::value,
-    void>::type
+std::enable_if_t<
+    std::is_same_v<data_t, at::Half> || std::is_same_v<data_t, at::BFloat16>,
+    void>
 index_select_add(
    const Tensor& select_indices,
    const Tensor& add_indices,
@ -366,7 +365,7 @@ index_select_add(
  }
 }
 template<typename data_t, typename index_t>
-typename std::enable_if<std::is_same<data_t, float>::value, void>::type
+std::enable_if_t<std::is_same_v<data_t, float>, void>
 index_select_add(const Tensor &select_indices,
                             const Tensor &add_indices,
                             const Tensor &src,
@ -493,7 +492,7 @@ index_select_add(const Tensor &select_indices,
 // mul (scaling by per_sample_weights)
 // index_add (using add_indices as the index)
 template <typename data_t, typename index_t>
-static typename std::enable_if<std::is_same<data_t, double>::value, void>::type
+static std::enable_if_t<std::is_same_v<data_t, double>, void>
 index_select_scale_add(
    const Tensor& select_indices,
    const Tensor& add_indices,
@ -548,10 +547,9 @@ index_select_scale_add(
 }

 template <typename data_t, typename index_t>
-typename std::enable_if<
-    std::is_same<data_t, at::Half>::value ||
-        std::is_same<data_t, at::BFloat16>::value,
-    void>::type
+std::enable_if_t<
+    std::is_same_v<data_t, at::Half> || std::is_same_v<data_t, at::BFloat16>,
+    void>
 index_select_scale_add(
    const Tensor& select_indices,
    const Tensor& add_indices,
@ -741,7 +739,7 @@ index_select_scale_add(
  }
 }
 template<typename data_t, typename index_t>
-typename std::enable_if<std::is_same<data_t, float>::value, void>::type
+std::enable_if_t<std::is_same_v<data_t, float>, void>
 index_select_scale_add(const Tensor &select_indices,
                                          const Tensor &add_indices,
                                          const Tensor &scale,
--- a/aten/src/ATen/native/IndexingUtils.h
+++ b/aten/src/ATen/native/IndexingUtils.h
@ -14,7 +14,7 @@ static void invalid_mask(const Tensor & self, int64_t idx, const Tensor & mask,
 }


-static C10_UNUSED std::vector<Tensor> expandTensors(const Tensor & self, IOptTensorListRef indices) {
+C10_UNUSED static std::vector<Tensor> expandTensors(const Tensor & self, IOptTensorListRef indices) {
  // If indices come in as ByteTensor or BoolTensor (masks), expand them into the equivalent indexing by LongTensors
  std::vector<Tensor> result;
  for (const auto& index_opt : indices) {
@ -48,7 +48,7 @@ static C10_UNUSED std::vector<Tensor> expandTensors(const Tensor & self, IOptTen
  return result;
 }

-static C10_UNUSED void checkIndexTensorTypes(IOptTensorListRef indices, bool allow_int=false) {
+C10_UNUSED static void checkIndexTensorTypes(IOptTensorListRef indices, bool allow_int=false) {
  for (const auto& tensor : indices) {
    if (tensor.has_value() && tensor->defined()) {
      auto scalarType = tensor->scalar_type();
@ -83,7 +83,7 @@ inline torch::List<std::optional<Tensor>> toListOfOptionalTensors(ArrayRef<IValu
  return result;
 }

-static C10_UNUSED bool hasContiguousSubspace(TensorList tl) {
+C10_UNUSED static bool hasContiguousSubspace(TensorList tl) {
  // true if all the non-null tensors are adjacent
  auto isDefined = [](const Tensor & tensor){ return tensor.defined(); };
  auto isNull = [](const Tensor & tensor){ return !tensor.defined(); };
@ -100,7 +100,7 @@ static C10_UNUSED bool hasContiguousSubspace(TensorList tl) {
 // transposeToFront(tensor, {nullptr, a, nullptr, b})
 // returns
 // tensor.permute([1, 3, 0, 2]), {a, b, nullptr, nullptr}
-static C10_UNUSED std::tuple<Tensor, std::vector<Tensor>>
+C10_UNUSED static std::tuple<Tensor, std::vector<Tensor>>
 transposeToFront(const Tensor& self, TensorList indices) {
  std::vector<int64_t> dims;
  std::vector<Tensor> transposedIndices;
--- a/aten/src/ATen/native/Math.h
+++ b/aten/src/ATen/native/Math.h
@ -3,6 +3,7 @@
 #include <ATen/AccumulateType.h>
 #include <ATen/NumericUtils.h>
 #include <ATen/jiterator_macros.h>
+#include <c10/macros/Macros.h>
 #include <c10/util/BFloat16.h>
 #include <c10/util/Half.h>
 #include <c10/util/MathConstants.h>
@ -3071,14 +3072,14 @@ inline C10_HOST_DEVICE T hermite_polynomial_h_forward(T x, int64_t n) {
    return r;
 } // hermite_polynomial_h_forward(T x, int64_t n)

-template<typename T, bool is_cuda=false, std::enable_if_t<!std::is_floating_point<T>::value, int> = 0>
+template<typename T, bool is_cuda=false, std::enable_if_t<!std::is_floating_point_v<T>, int> = 0>
 inline C10_HOST_DEVICE T hermite_polynomial_h_forward(T x, T n) {
    return hermite_polynomial_h_forward(x, static_cast<int64_t>(n));
 } // hermite_polynomial_h_forward(T x, T n)

-template<typename T, bool is_cuda=false, std::enable_if_t<std::is_floating_point<T>::value, int> = 0>
-inline C10_HOST_DEVICE T hermite_polynomial_h_forward(T x, T n) {
-    return hermite_polynomial_h_forward(x, ((!std::isinf(n)) && (!std::isnan(n))) ? static_cast<int64_t>(n) : static_cast<int64_t>(-1));
+template<typename T, bool is_cuda=false, std::enable_if_t<std::is_floating_point_v<T>, int> = 0>
+__ubsan_ignore_float_cast_overflow__ inline C10_HOST_DEVICE T hermite_polynomial_h_forward(T x, T n) {
+    return hermite_polynomial_h_forward(x, (!std::isinf(n) && !std::isnan(n)) ? static_cast<int64_t>(n) : static_cast<int64_t>(-1));
 } // hermite_polynomial_h_forward(T x, T n)

 template<typename T>
--- a/aten/src/ATen/native/Pow.h
+++ b/aten/src/ATen/native/Pow.h
@ -23,7 +23,7 @@ namespace native {
 // e.g. since 2**-1==0.5, the truncated integral result is zero. 1**negative_exponent is the
 // only non-zero result.
 template <class T,
-  typename std::enable_if<std::is_integral<T>::value, T>::type* = nullptr>
+  std::enable_if_t<std::is_integral_v<T>, T>* = nullptr>
 inline HOST_DEVICE __ubsan_ignore_signed_int_overflow__ T powi_impl(T a, T b) {
  T result = 1;
  while (b) {
@ -37,13 +37,13 @@ inline HOST_DEVICE __ubsan_ignore_signed_int_overflow__ T powi_impl(T a, T b) {
 }

 template <class T,
-  typename std::enable_if<std::is_integral<T>::value && !std::is_signed<T>::value, T>::type* = nullptr>
+  std::enable_if_t<std::is_integral_v<T> && !std::is_signed_v<T>, T>* = nullptr>
 inline HOST_DEVICE T powi(T a, T b) {
  return powi_impl(a, b);
 }

 template <class T,
-  typename std::enable_if<std::is_integral<T>::value && std::is_signed<T>::value, T>::type* = nullptr>
+  std::enable_if_t<std::is_integral_v<T> && std::is_signed_v<T>, T>* = nullptr>
 inline HOST_DEVICE T powi(T a, T b) {
  if ( b < 0 ) {
      if ( a == 1 ) {
--- a/aten/src/ATen/native/RNN.cpp
+++ b/aten/src/ATen/native/RNN.cpp
@ -179,7 +179,7 @@ struct CellParams : public CellParamsBase {
      const Tensor& _b_ih,
      const Tensor& _b_hh,
      const Tensor& _w_hr)
-      : w_ih(_w_ih), w_hh(_w_hh), b_ih_(_b_ih), b_hh_(_b_hh), w_hr(_w_hr) {};
+      : w_ih(_w_ih), w_hh(_w_hh), b_ih_(_b_ih), b_hh_(_b_hh), w_hr(_w_hr) {}

  const Tensor& w_ih;
  const Tensor& w_hh;
@ -825,7 +825,7 @@ struct FullLayer : Layer<Tensor, hidden_type, cell_params> {
  using unstacked_output_type = LayerOutput<std::vector<Tensor>, hidden_type>;

  FullLayer(Cell<hidden_type, cell_params>& cell)
-    : cell_(cell) {};
+    : cell_(cell) {}

  unstacked_output_type operator()(
      const std::vector<Tensor>& step_inputs,
@ -870,7 +870,7 @@ struct FullBidirectionalLayer
  using output_type = typename Layer<Tensor, hidden_type, param_type>::output_type;

  FullBidirectionalLayer(Cell<dir_hidden_type, cell_params>& cell)
-    : layer_(cell) {};
+    : layer_(cell) {}

  output_type operator()(
      const Tensor& input,
@ -922,7 +922,7 @@ struct PackedLayer : Layer<PackedSequence, hidden_type, cell_params> {
      typename Layer<PackedSequence, hidden_type, cell_params>::output_type;

  PackedLayer(Cell<hidden_type, cell_params>& cell)
-    : cell_(cell) {};
+    : cell_(cell) {}

  output_type operator()(
      const PackedSequence& input,
@ -983,7 +983,7 @@ struct ReversedPackedLayer : Layer<PackedSequence, hidden_type, cell_params> {
      typename Layer<PackedSequence, hidden_type, cell_params>::output_type;

  ReversedPackedLayer(Cell<hidden_type, cell_params>& cell)
-    : cell_(cell) {};
+    : cell_(cell) {}

  output_type operator()(
      const PackedSequence& input,
@ -1040,7 +1040,7 @@ struct PackedBidirectionalLayer
      typename Layer<PackedSequence, hidden_type, param_type>::output_type;

  PackedBidirectionalLayer(Cell<dir_hidden_type, cell_params>& cell)
-    : layer_(cell), rev_layer_(cell) {};
+    : layer_(cell), rev_layer_(cell) {}

  output_type operator()(
      const PackedSequence& input,
@ -1889,7 +1889,7 @@ static DEFINE_QUANTIZED_RNN_CELL_DYNAMIC(quantized_rnn_tanh_cell_dynamic, simple

 namespace {

-static C10_UNUSED auto ensure_linear_params_registered = register_linear_params();
+C10_UNUSED static auto ensure_linear_params_registered = register_linear_params();

 static auto cell_params_base_registry =
    torch::selective_class_<CellParamsBase>("rnn", TORCH_SELECTIVE_CLASS("CellParamsBase"))
--- a/aten/src/ATen/native/ReduceOps.cpp
+++ b/aten/src/ATen/native/ReduceOps.cpp
@ -753,11 +753,11 @@ Tensor cumprod_backward(const Tensor& grad, const Tensor& input, int64_t dim, co
 namespace {
 #ifdef _MSC_VER
 template<typename T>
-inline typename std::enable_if<std::is_integral<T>::value, bool>::type isnan_(T x) {
+inline std::enable_if_t<std::is_integral_v<T>, bool> isnan_(T x) {
  return false;
 }
 template<typename T>
-inline typename std::enable_if<!std::is_integral<T>::value, bool>::type isnan_(T x) {
+inline std::enable_if_t<!std::is_integral_v<T>, bool> isnan_(T x) {
  return std::isnan(x);
 }
 #else
--- a/aten/src/ATen/native/ReduceOpsUtils.h
+++ b/aten/src/ATen/native/ReduceOpsUtils.h
@ -207,7 +207,7 @@ inline TensorIterator make_reduction(
  return TensorIterator::reduce_op(viewed_result, self.to(in_dtype));
 }

-inline C10_UNUSED TensorIterator make_reduction(
+C10_UNUSED inline TensorIterator make_reduction(
    const char* name, Tensor& result, const Tensor& self,
    at::OptionalIntArrayRef dim, bool keepdim, ScalarType out_dtype) {
  // special case for type promotion in mixed precision, improves computational
@ -259,7 +259,7 @@ inline TensorIterator make_reduction(
  return TensorIterator::reduce_op(viewed_result1, viewed_result2, self.to(dtype1));
 }

-inline C10_UNUSED TensorIterator make_reduction(
+C10_UNUSED inline TensorIterator make_reduction(
    const char* name, Tensor& result1, Tensor& result2, const Tensor& self,
    at::OptionalIntArrayRef dim, bool keepdim, ScalarType dtype) {
  return make_reduction(name, result1, result2, self, dim, keepdim, dtype, dtype);
@ -313,7 +313,7 @@ inline std::vector<int64_t> get_zero_numel_tensor_size(
 // This function should be called when you are reducing a zero-numel tensor and want to
 // resize the output and return it. This function exists for resizing zero-numel
 // tensors when the size of the reduction dimension is non-zero.
-inline C10_UNUSED void zero_numel_tensor_resize(Tensor& result, Tensor& result_indices,
+C10_UNUSED inline void zero_numel_tensor_resize(Tensor& result, Tensor& result_indices,
                                     const Tensor& self, const int64_t dim,
                                     const bool keepdim, const char *fn_name) {
  auto sizes = get_zero_numel_tensor_size(self, dim, keepdim, fn_name);
@ -349,7 +349,7 @@ inline ScalarType get_dtype_from_result(Tensor& result, std::optional<ScalarType

 namespace at::meta {

-inline C10_UNUSED DimVector get_reduction_shape(
+C10_UNUSED inline DimVector get_reduction_shape(
    const Tensor& self,
    IntArrayRef dims,
    bool keepdim,
@ -434,7 +434,7 @@ inline TensorIterator make_reduction(
  return TensorIterator::reduce_op(viewed_result1, viewed_result2, self.to(dtype1));
 }

-inline C10_UNUSED TensorIterator make_reduction_from_out_ty(
+C10_UNUSED inline TensorIterator make_reduction_from_out_ty(
    const Tensor& self,
    const Tensor& result,
    OptionalIntArrayRef opt_dims,
--- a/aten/src/ATen/native/ScatterGatherChecks.h
+++ b/aten/src/ATen/native/ScatterGatherChecks.h
@ -52,7 +52,7 @@ inline void gather_shape_check(const Tensor& self, int64_t dim,
        ensure_nonempty_size(index, i) <= ensure_nonempty_size(self, i),
        "Size does not match at dimension ", i,
        " expected index ", index.sizes(),
-        " to be smaller than self ", self.sizes(),
+        " to be no larger than self ", self.sizes(),
        " apart from dimension ", dim
      );
    }
@ -109,15 +109,15 @@ inline void scatter_shape_check(

    TORCH_CHECK(!is_wrong_shape,
      "Expected index ", index.sizes(),
-      " to be smaller than self ", self.sizes(),
+      " to be no larger than self ", self.sizes(),
      " apart from dimension ", dim,
-      " and to be smaller size than src ", src.sizes()
+      " and to be no larger size than src ", src.sizes()
    );
  }
  else {
    TORCH_CHECK(!is_wrong_shape,
      "Expected index ", index.sizes(),
-      " to be smaller than self ", self.sizes(),
+      " to be no larger than self ", self.sizes(),
      " apart from dimension ", dim
    );
  }
--- a/aten/src/ATen/native/TensorShape.cpp
+++ b/aten/src/ATen/native/TensorShape.cpp
@ -216,15 +216,6 @@
 #include <vector>

 namespace at::meta {
-inline void cat_check_no_zero_dim(const MaterializedITensorListRef& tensors) {
-  size_t i = 0;
-  for (const Tensor& t : tensors) {
-    TORCH_CHECK(
-        t.dim() > 0,
-        "zero-dimensional tensor (at position ", i, ") cannot be concatenated");
-    i++;
-  }
-}

 inline c10::MemoryFormat cat_compute_output_memory_format(const MaterializedITensorListRef& inputs) {
  std::optional<c10::MemoryFormat> format = std::nullopt;
@ -248,7 +239,7 @@ TORCH_PRECOMPUTE_META_FUNC(cat)(const ITensorListRef& tensors, int64_t dim) {
  // size (i.e. other empty sizes are not skipped).
  auto materialized = tensors.materialize();

-  cat_check_no_zero_dim(materialized);
+  native::check_cat_no_zero_dim(materialized);
  dim = at::legacy_cat_wrap_dim(dim, materialized);

  // Checking names before the actual dimensions.
--- a/aten/src/ATen/native/TensorShape.h
+++ b/aten/src/ATen/native/TensorShape.h
@ -30,7 +30,7 @@ inline void check_cat_shape_except_dim(const Tensor & first, const Tensor & seco
 }

 inline void check_cat_no_zero_dim(const MaterializedITensorListRef& tensors) {
-  int64_t i = 0;
+  [[maybe_unused]] int64_t i = 0;
  for(const Tensor& t : tensors) {
    TORCH_CHECK(t.dim() > 0,
             "zero-dimensional tensor (at position ", i, ") cannot be concatenated");
--- a/aten/src/ATen/native/UnfoldBackward.h
+++ b/aten/src/ATen/native/UnfoldBackward.h
@ -29,7 +29,7 @@ namespace {
 // grad_in does not mean that it is a gradient wrt to input,
 // grad_in/grad_out is just an input/output of unfold_backward kernel.

-static C10_UNUSED TensorIterator _make_unfold_backward_iter_over_grad_out(
+C10_UNUSED static TensorIterator _make_unfold_backward_iter_over_grad_out(
  Tensor& grad_out,
  const Tensor& grad_in,
  int64_t dim,
--- a/aten/src/ATen/native/UpSample.h
+++ b/aten/src/ATen/native/UpSample.h
@ -103,7 +103,7 @@ DECLARE_DISPATCH(upsampling_bicubic2d, upsample_bicubic2d_kernel);
 DECLARE_DISPATCH(_upsampling_bicubic2d_aa, _upsample_bicubic2d_aa_kernel);
 DECLARE_DISPATCH(_upsampling_bicubic2d_aa, _upsample_bicubic2d_aa_backward_kernel);

-inline C10_UNUSED std::array<int64_t, 3> upsample_1d_common_check(IntArrayRef input_size, IntArrayRef output_size) {
+C10_UNUSED inline std::array<int64_t, 3> upsample_1d_common_check(IntArrayRef input_size, IntArrayRef output_size) {
  TORCH_CHECK(
      output_size.size() == 1,
      "It is expected output_size equals to 1, but got size ",
@ -131,7 +131,7 @@ inline C10_UNUSED std::array<int64_t, 3> upsample_1d_common_check(IntArrayRef in
  return {nbatch, channels, output_width};
 }

-inline C10_UNUSED std::array<int64_t, 4> upsample_2d_common_check(IntArrayRef input_size, IntArrayRef output_size) {
+C10_UNUSED inline std::array<int64_t, 4> upsample_2d_common_check(IntArrayRef input_size, IntArrayRef output_size) {
  TORCH_CHECK(
      output_size.size() == 2,
      "It is expected output_size equals to 2, but got size ",
@ -167,7 +167,7 @@ inline C10_UNUSED std::array<int64_t, 4> upsample_2d_common_check(IntArrayRef in
  return {nbatch, channels, output_height, output_width};
 }

-inline C10_UNUSED
+C10_UNUSED inline
 std::array<int64_t, 5> upsample_3d_common_check(IntArrayRef input_size, IntArrayRef output_size) {
  TORCH_CHECK(
      output_size.size() == 3,
--- a/aten/src/ATen/native/ao_sparse/quantized/cpu/fbgemm_utils.cpp
+++ b/aten/src/ATen/native/ao_sparse/quantized/cpu/fbgemm_utils.cpp
@ -40,7 +40,7 @@ int register_linear_params() {
 }

 namespace {
-static C10_UNUSED auto linear_params = register_linear_params();
+C10_UNUSED static auto linear_params = register_linear_params();
 }  // namespace

 }}  // namespace ao::sparse
--- a/aten/src/ATen/native/cpu/BlasKernel.cpp
+++ b/aten/src/ATen/native/cpu/BlasKernel.cpp
@ -96,7 +96,7 @@ auto sum(int64_t N, Func f) {
 }

 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<std::is_same<scalar_t, opmath_t>::value, void>::type
+std::enable_if_t<std::is_same_v<scalar_t, opmath_t>, void>
 gemm_notrans_(
    int64_t m,
    int64_t n,
@ -132,7 +132,7 @@ gemm_notrans_(

 // std::is_same<scalar_t, at::BFloat16> || std::is_same<scalar_t, at::Half>
 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<!std::is_same<scalar_t, opmath_t>::value, void>::type
+std::enable_if_t<!std::is_same_v<scalar_t, opmath_t>, void>
 gemm_notrans_(
    int64_t m,
    int64_t n,
@ -222,7 +222,7 @@ void gemm_transb_impl(
 }

 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<std::is_same<scalar_t, opmath_t>::value, void>::type
+std::enable_if_t<std::is_same_v<scalar_t, opmath_t>, void>
 gemm_transb_(
    TransposeType transb,
    int64_t m,
@ -244,7 +244,7 @@ gemm_transb_(

 // std::is_same<scalar_t, at::BFloat16> || std::is_same<scalar_t, at::Half>
 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<!std::is_same<scalar_t, opmath_t>::value, void>::type
+std::enable_if_t<!std::is_same_v<scalar_t, opmath_t>, void>
 gemm_transb_(
    TransposeType transb,
    int64_t m,
--- a/aten/src/ATen/native/cpu/CopyKernel.cpp
+++ b/aten/src/ATen/native/cpu/CopyKernel.cpp
@ -82,7 +82,7 @@ static void reduced_float_copy_kernel(TensorIteratorBase &iter, bool requires_ne
        std::copy_n(base, 2, data.data());
        const int64_t *outer_strides = &strides[2];

-        for (const auto it C10_UNUSED : c10::irange(size1)) {
+        for (C10_UNUSED const auto it : c10::irange(size1)) {
          Vecd dst_s;
          if (strides_in[0] == 0) {
            dst_s = Vecd(dest_t(*((scalar_t*)data[1])));
@ -151,7 +151,7 @@ static void reduced_float_copy_kernel(TensorIteratorBase &iter, bool requires_ne
        std::copy_n(base, 2, data.data());
        const int64_t *outer_strides = &strides[2];

-        for (const auto it C10_UNUSED : c10::irange(size1)) {
+        for (C10_UNUSED const auto it : c10::irange(size1)) {
          Vecd dst_s;
          if (strides_in[0] == 0) {
            dst_s = Vecd(dest_t(*((source_t*)data[1])));
--- a/aten/src/ATen/native/cpu/FusedAdagradKernel.cpp
+++ b/aten/src/ATen/native/cpu/FusedAdagradKernel.cpp
@ -12,10 +12,10 @@ namespace at::native {
 namespace{

 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<
-    std::is_same<scalar_t, Half>::value || std::is_same<scalar_t, BFloat16>::value,
-    void>::
-    type inline adagrad_math(
+std::enable_if_t<
+    std::is_same_v<scalar_t, Half> || std::is_same_v<scalar_t, BFloat16>,
+    void>
+    inline adagrad_math(
  scalar_t* param_ptr,
  scalar_t* grad_ptr,
  scalar_t* state_sum_ptr,
@ -81,10 +81,10 @@ typename std::enable_if<


 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<
-    std::is_same<scalar_t, float>::value || std::is_same<scalar_t, double>::value,
-    void>::
-    type inline adagrad_math(
+std::enable_if_t<
+    std::is_same_v<scalar_t, float> || std::is_same_v<scalar_t, double>,
+    void>
+    inline adagrad_math(
  scalar_t* param_ptr,
  scalar_t* grad_ptr,
  scalar_t* state_sum_ptr,
--- a/aten/src/ATen/native/cpu/FusedAdamKernel.cpp
+++ b/aten/src/ATen/native/cpu/FusedAdamKernel.cpp
@ -12,10 +12,10 @@ namespace at::native {
 namespace{

 template <typename scalar_t, typename opmath_t, ADAM_MODE adam_mode>
-typename std::enable_if<
-    std::is_same<scalar_t, Half>::value || std::is_same<scalar_t, BFloat16>::value,
-    void>::
-    type inline adam_math(
+std::enable_if_t<
+    std::is_same_v<scalar_t, Half> || std::is_same_v<scalar_t, BFloat16>,
+    void>
+    inline adam_math(
  scalar_t* param_ptr,
  scalar_t* exp_avg_ptr,
  scalar_t* exp_avg_sq_ptr,
@ -155,10 +155,10 @@ typename std::enable_if<


 template <typename scalar_t, typename opmath_t, ADAM_MODE adam_mode>
-typename std::enable_if<
-    std::is_same<scalar_t, float>::value || std::is_same<scalar_t, double>::value,
-    void>::
-    type inline adam_math(
+std::enable_if_t<
+    std::is_same_v<scalar_t, float> || std::is_same_v<scalar_t, double>,
+    void>
+    inline adam_math(
  scalar_t* param_ptr,
  scalar_t* exp_avg_ptr,
  scalar_t* exp_avg_sq_ptr,
--- a/aten/src/ATen/native/cpu/FusedSGDKernel.cpp
+++ b/aten/src/ATen/native/cpu/FusedSGDKernel.cpp
@ -12,10 +12,10 @@ namespace at::native {
 namespace{

 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<
-    std::is_same<scalar_t, Half>::value || std::is_same<scalar_t, BFloat16>::value,
-    void>::
-    type inline sgd_math(
+std::enable_if_t<
+    std::is_same_v<scalar_t, Half> || std::is_same_v<scalar_t, BFloat16>,
+    void>
+    inline sgd_math(
  scalar_t* param_ptr,
  scalar_t* grad_ptr,
  scalar_t* momentum_buf_ptr,
@ -104,10 +104,10 @@ typename std::enable_if<


 template <typename scalar_t, typename opmath_t>
-typename std::enable_if<
-    std::is_same<scalar_t, float>::value || std::is_same<scalar_t, double>::value,
-    void>::
-    type inline sgd_math(
+std::enable_if_t<
+    std::is_same_v<scalar_t, float> || std::is_same_v<scalar_t, double>,
+    void>
+    inline sgd_math(
  scalar_t* param_ptr,
  scalar_t* grad_ptr,
  scalar_t* momentum_buf_ptr,
--- a/aten/src/ATen/native/cpu/IsContiguous.h
+++ b/aten/src/ATen/native/cpu/IsContiguous.h
@ -31,14 +31,16 @@ struct IsContiguous<0, -1, traits, s> {
 };

 // output and all inputs are contiguous
-template <typename traits,
-    typename std::enable_if<std::is_void<typename traits::result_type>::value>::type* = nullptr>
+template <
+    typename traits,
+    std::enable_if_t<std::is_void_v<typename traits::result_type>>* =
+        nullptr>
 static inline bool is_contiguous(const int64_t* strides) {
  return IsContiguous<traits::arity, traits::arity - 1, traits>::eval(strides);
 }

 template <typename traits,
-    typename std::enable_if<!std::is_void<typename traits::result_type>::value>::type* = nullptr>
+    std::enable_if_t<!std::is_void_v<typename traits::result_type>>* = nullptr>
 static inline bool is_contiguous(const int64_t* strides) {
  return IsContiguous<traits::arity, traits::arity, traits>::eval(strides);
 }
@ -46,14 +48,14 @@ static inline bool is_contiguous(const int64_t* strides) {
 // input at `s` is scalar (stride 0); output and other inputs are contiguous
 // NB: output is typically at strides[0] so first input corresponds to s=1
 template <typename traits, int s,
-    typename std::enable_if<std::is_void<typename traits::result_type>::value>::type* = nullptr>
+    std::enable_if_t<std::is_void_v<typename traits::result_type>>* = nullptr>
 static inline bool is_contiguous_scalar(const int64_t* strides) {
  static_assert(s > 0 && s <= traits::arity, "scalar argument index out of bounds");
  return IsContiguous<traits::arity, traits::arity - 1, traits, s>::eval(strides);
 }

 template <typename traits, int s,
-    typename std::enable_if<!std::is_void<typename traits::result_type>::value>::type* = nullptr>
+    std::enable_if_t<!std::is_void_v<typename traits::result_type>>* = nullptr>
 static inline bool is_contiguous_scalar(const int64_t* strides) {
  static_assert(s > 0 && s <= traits::arity, "scalar argument index out of bounds");
  return IsContiguous<traits::arity, traits::arity, traits, s>::eval(strides);
--- a/aten/src/ATen/native/cpu/Loops.h
+++ b/aten/src/ATen/native/cpu/Loops.h
@ -271,7 +271,7 @@ struct VectorizedLoop2d {
    const int64_t *outer_strides = &strides[ntensors];

    if (is_contiguous<traits>(strides)) {
-      for (const auto i C10_UNUSED : c10::irange(size1)) {
+      for (C10_UNUSED const auto i : c10::irange(size1)) {
        vectorized_loop(data.data(), size0, 0, op, vop);
        advance(data, outer_strides);
      }
@ -279,12 +279,12 @@ struct VectorizedLoop2d {
      using Indices = std::make_index_sequence<traits::arity>;
      unroll_contiguous_scalar_checks<traits>(strides, Indices{}, [&](size_t idx) {
        if (idx) {
-          for (const auto i C10_UNUSED : c10::irange(size1)) {
+          for (C10_UNUSED const auto i : c10::irange(size1)) {
            vectorized_loop(data.data(), size0, idx, op, vop);
            advance(data, outer_strides);
          }
        } else {
-          for (const auto i C10_UNUSED : c10::irange(size1)) {
+          for (C10_UNUSED const auto i : c10::irange(size1)) {
            basic_loop(data.data(), strides, 0, size0, op);
            advance(data, outer_strides);
          }
--- a/aten/src/ATen/native/cpu/MaxPoolKernel.cpp
+++ b/aten/src/ATen/native/cpu/MaxPoolKernel.cpp
@ -64,7 +64,7 @@ vec::Vectorized<int64_t> is_nan_vec<int64_t>(vec::Vectorized<int64_t> vec) {

 template <typename scalar_t, typename opmath_t>
 inline
-typename std::enable_if<std::is_same<scalar_t, opmath_t>::value, void>::type
+std::enable_if_t<std::is_same_v<scalar_t, opmath_t>, void>
 compute_internal(
  const scalar_t* input_data,
  scalar_t* out_data,
@ -139,7 +139,7 @@ compute_internal(
 // std::is_same<scalar_t, at::BFloat16> || std::is_same<scalar_t, at::Half>
 template <typename scalar_t, typename opmath_t>
 inline
-typename std::enable_if<!std::is_same<scalar_t, opmath_t>::value, void>::type
+std::enable_if_t<!std::is_same_v<scalar_t, opmath_t>, void>
 compute_internal(
  const scalar_t* input_data,
  scalar_t* out_data,
--- a/aten/src/ATen/native/cpu/Reduce.h
+++ b/aten/src/ATen/native/cpu/Reduce.h
@ -129,13 +129,13 @@ static void set_results(const res_t result, const TensorIteratorBase &iter, cons
 }

 template<typename traits, std::size_t i = 0, typename... tuple_t>
-inline typename std::enable_if<i == sizeof...(tuple_t), std::size_t>::type
+inline std::enable_if_t<i == sizeof...(tuple_t), std::size_t>
 for_each_in_tuple(const std::tuple<tuple_t...>& /*t*/, const TensorIteratorBase& /*iter*/, const int /*num_outputs*/) {
  return i;
 }

 template<typename traits, std::size_t i = 0, typename... tuple_t>
-inline typename std::enable_if<i < sizeof...(tuple_t), std::size_t>::type
+inline std::enable_if_t<i < sizeof...(tuple_t), std::size_t>
 for_each_in_tuple(const std::tuple<tuple_t...>& t, const TensorIteratorBase &iter, const int num_outputs) {
  if (i < (size_t)num_outputs) {
    set_result<traits>(i, std::get<i>(t), iter, num_outputs);
--- a/aten/src/ATen/native/cpu/ReduceUtils.h
+++ b/aten/src/ATen/native/cpu/ReduceUtils.h
@ -106,7 +106,7 @@ inline void _init(scalar_t* self_ptr, at::opmath_type<scalar_t>* buffer_ptr, int
 }

 template <typename scalar_t>
-inline typename std::enable_if<!std::is_same<scalar_t, Vec2>::value, scalar_t>::type
+inline std::enable_if_t<!std::is_same_v<scalar_t, Vec2>, scalar_t>
 _max(const scalar_t& x, const scalar_t& y) {
  return at::_isnan(y) ? y : std::max(x, y);
 }
@ -118,14 +118,14 @@ inline Vectorized<scalar_t> _max(const Vectorized<scalar_t>& x, const Vectorized
 }

 template <typename vec_t>
-inline typename std::enable_if<std::is_same<vec_t, Vec2>::value, Vec2>::type
+inline std::enable_if_t<std::is_same_v<vec_t, Vec2>, Vec2>
 _max(const vec_t& x, const vec_t& y) {
  // vec::maximum propagates NaN
  return maximum(x, y);
 }

 template <typename scalar_t>
-inline typename std::enable_if<!std::is_same<scalar_t, Vec2>::value, scalar_t>::type
+inline std::enable_if_t<!std::is_same_v<scalar_t, Vec2>, scalar_t>
 _min(const scalar_t& x, const scalar_t& y) {
  return at::_isnan(y) ? y : std::min(x, y);
 }
@ -137,7 +137,7 @@ inline Vectorized<scalar_t> _min(const Vectorized<scalar_t>& x, const Vectorized
 }

 template <typename vec_t>
-inline typename std::enable_if<std::is_same<vec_t, Vec2>::value, Vec2>::type
+inline std::enable_if_t<std::is_same_v<vec_t, Vec2>, Vec2>
 _min(const vec_t& x, const vec_t& y) {
  // vec::minimum propagates NaN
  return minimum(x, y);
--- a/aten/src/ATen/native/cpu/UnfoldBackwardKernel.cpp
+++ b/aten/src/ATen/native/cpu/UnfoldBackwardKernel.cpp
@ -76,7 +76,7 @@ void _unfold_backward_internal_kernel(
    auto* RESTRICT grad_in_ptr = data[1];
    auto* RESTRICT idx_dim_ptr = data[2];

-    for (const auto elem C10_UNUSED : c10::irange(nelems)) {
+    for (C10_UNUSED const auto elem : c10::irange(nelems)) {
      auto* RESTRICT grad_out_data = reinterpret_cast<scalar_t*>(grad_out_ptr);
      auto* RESTRICT grad_in_data = reinterpret_cast<scalar_t*>(grad_in_ptr);

--- a/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
+++ b/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
@ -102,7 +102,7 @@ void pack_rgb(

  TORCH_INTERNAL_ASSERT(unpacked_increment == 3 || unpacked_increment == 4);

-  for (const auto i C10_UNUSED : c10::irange(num_pixels)) {
+  for (C10_UNUSED const auto i : c10::irange(num_pixels)) {
    for (const auto j : c10::irange(num_channels)) {
      packed[j * packed_stride] = unpacked[j];
    }
--- a/aten/src/ATen/native/cpu/group_norm_kernel.cpp
+++ b/aten/src/ATen/native/cpu/group_norm_kernel.cpp
@ -85,8 +85,8 @@ void GroupNormKernelImplInternal(
 }

 template <typename T>
-typename std::enable_if<std::is_same<T, at::opmath_type<T>>::value,
-  std::tuple<T, T>>::type
+std::enable_if_t<std::is_same_v<T, at::opmath_type<T>>,
+  std::tuple<T, T>>
 ColumnwiseMoments(
    const T* X_data,
    int64_t HxW,
@ -118,8 +118,8 @@ ColumnwiseMoments(

 // std::is_same<T, at::BFloat16> || std::is_same<T, at::Half>
 template <typename T>
-typename std::enable_if<!std::is_same<T, at::opmath_type<T>>::value,
-  std::tuple<at::opmath_type<T>, at::opmath_type<T>>>::type
+std::enable_if_t<!std::is_same_v<T, at::opmath_type<T>>,
+  std::tuple<at::opmath_type<T>, at::opmath_type<T>>>
 ColumnwiseMoments(
    const T* X_data,
    int64_t HxW,
@ -160,7 +160,7 @@ ColumnwiseMoments(
 }

 template <typename T, typename opmath_t>
-inline typename std::enable_if<std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<std::is_same_v<T, opmath_t>, void>
 CalcMeanVar(
  const T* X_ptr,
  opmath_t* mean_ptr,
@ -183,7 +183,7 @@ CalcMeanVar(

 // std::is_same<T, at::BFloat16> || std::is_same<T, at::Half>
 template <typename T, typename opmath_t>
-inline typename std::enable_if<!std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<!std::is_same_v<T, opmath_t>, void>
 CalcMeanVar(
  const T* X_ptr,
  opmath_t* mean_ptr,
@ -227,7 +227,7 @@ CalcMeanVar(
 }

 template <typename T, typename opmath_t>
-inline typename std::enable_if<std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<std::is_same_v<T, opmath_t>, void>
 ApplyScaleBias(
  T* Y_ptr,
  const T* X_ptr,
@ -246,7 +246,7 @@ ApplyScaleBias(

 // std::is_same<T, at::BFloat16> || std::is_same<T, at::Half>
 template <typename T, typename opmath_t>
-inline typename std::enable_if<!std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<!std::is_same_v<T, opmath_t>, void>
 ApplyScaleBias(
  T* Y_ptr,
  const T* X_ptr,
@ -529,7 +529,7 @@ void GroupNormKernelImpl(


 template <typename T, typename opmath_t>
-typename std::enable_if<std::is_same<T, opmath_t>::value, void>::type
+std::enable_if_t<std::is_same_v<T, opmath_t>, void>
 ComputeInternalGradients(
    int64_t N,
    int64_t C,
@ -556,7 +556,7 @@ ComputeInternalGradients(
 }

 template <typename T, typename opmath_t>
-typename std::enable_if<!std::is_same<T, opmath_t>::value, void>::type
+std::enable_if_t<!std::is_same_v<T, opmath_t>, void>
 ComputeInternalGradients(
    int64_t N,
    int64_t C,
@ -603,7 +603,7 @@ ComputeInternalGradients(
 }

 template <typename PT, typename opmath_t>
-inline typename std::enable_if<std::is_same<PT, opmath_t>::value, void>::type
+inline std::enable_if_t<std::is_same_v<PT, opmath_t>, void>
 CalcDsDb(
    const opmath_t* ds_ptr,
    const opmath_t* db_ptr,
@ -626,7 +626,7 @@ CalcDsDb(
 }

 template <typename PT, typename opmath_t>
-inline typename std::enable_if<!std::is_same<PT, opmath_t>::value, void>::type
+inline std::enable_if_t<!std::is_same_v<PT, opmath_t>, void>
 CalcDsDb(
    const opmath_t* ds_ptr,
    const opmath_t* db_ptr,
@ -708,7 +708,7 @@ void GroupNormInputBackward(
 }

 template <typename PT, typename opmath_t>
-typename std::enable_if<std::is_same<PT, opmath_t>::value, void>::type
+std::enable_if_t<std::is_same_v<PT, opmath_t>, void>
 GammaBackward(
    int64_t N,
    int64_t C,
@ -755,7 +755,7 @@ GammaBackward(
 }

 template <typename PT, typename opmath_t>
-typename std::enable_if<!std::is_same<PT, opmath_t>::value, void>::type
+std::enable_if_t<!std::is_same_v<PT, opmath_t>, void>
 GammaBackward(
    int64_t N,
    int64_t C,
@ -817,7 +817,7 @@ GammaBackward(
 }

 template <typename PT, typename opmath_t>
-typename std::enable_if<std::is_same<PT, opmath_t>::value, void>::type
+std::enable_if_t<std::is_same_v<PT, opmath_t>, void>
 BetaBackward(int64_t N, int64_t C, const opmath_t* db, PT* dbeta) {
  using Vec = at::vec::Vectorized<PT>;
  constexpr int64_t K = Vec::size();
@ -841,7 +841,7 @@ BetaBackward(int64_t N, int64_t C, const opmath_t* db, PT* dbeta) {
 }

 template <typename PT, typename opmath_t>
-typename std::enable_if<!std::is_same<PT, opmath_t>::value, void>::type
+std::enable_if_t<!std::is_same_v<PT, opmath_t>, void>
 BetaBackward(int64_t N, int64_t C, const opmath_t* db, PT* dbeta) {
  using Vec = at::vec::Vectorized<PT>;
  using fVec = at::vec::Vectorized<opmath_t>;
@ -937,7 +937,7 @@ void GroupNormBackwardKernelImplInternal(
 }

 template <typename T, typename opmath_t>
-inline typename std::enable_if<std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<std::is_same_v<T, opmath_t>, void>
 DsDbRowwiseMomentsChannelsLast(
  const T* dY_ptr,
  const T* X_ptr,
@ -972,7 +972,7 @@ DsDbRowwiseMomentsChannelsLast(
 }

 template <typename T, typename opmath_t>
-inline typename std::enable_if<!std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<!std::is_same_v<T, opmath_t>, void>
 DsDbRowwiseMomentsChannelsLast(
  const T* dY_ptr,
  const T* X_ptr,
@ -1024,10 +1024,10 @@ DsDbRowwiseMomentsChannelsLast(
 }

 template <typename T>
-inline typename std::enable_if<std::is_same<T, at::opmath_type<T>>::value,
+inline std::enable_if_t<std::is_same_v<T, at::opmath_type<T>>,
  std::tuple<
  vec::Vectorized<T>,
-  vec::Vectorized<T>>>::type
+  vec::Vectorized<T>>>
 load_util(const T* data_ptr, int64_t n) {
  using Vec = vec::Vectorized<T>;
  auto vec0 = Vec::loadu(data_ptr, n > Vec::size() ? Vec::size() : n);
@ -1037,11 +1037,11 @@ load_util(const T* data_ptr, int64_t n) {
 }

 template <typename T>
-inline typename std::enable_if<!std::is_same<T, at::opmath_type<T>>::value,
+inline std::enable_if_t<!std::is_same_v<T, at::opmath_type<T>>,
  std::tuple<
    vec::Vectorized<at::opmath_type<T>>,
    vec::Vectorized<at::opmath_type<T>>>
-    >::type
+    >
 load_util(const T* data_ptr, int64_t n) {
  using Vec = vec::Vectorized<T>;
  auto vec = Vec::loadu(data_ptr, n);
@ -1049,7 +1049,7 @@ load_util(const T* data_ptr, int64_t n) {
 }

 template <typename T, typename PT, typename opmath_t>
-inline typename std::enable_if<std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<std::is_same_v<T, opmath_t>, void>
 ApplyInputGradientsChannelsLastColMov(
  const T* dY_data,
  const T* X_data,
@ -1097,7 +1097,7 @@ ApplyInputGradientsChannelsLastColMov(
 }

 template <typename T, typename PT, typename opmath_t>
-inline typename std::enable_if<!std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<!std::is_same_v<T, opmath_t>, void>
 ApplyInputGradientsChannelsLastColMov(
    const T* dY_data,
    const T* X_data,
@ -1154,7 +1154,7 @@ ApplyInputGradientsChannelsLastColMov(
 }

 template <typename T, typename PT, typename opmath_t>
-inline typename std::enable_if<std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<std::is_same_v<T, opmath_t>, void>
 ApplyInputGradientsChannelsLastRowMov(
  const T* dY_data,
  const T* X_data,
@ -1190,7 +1190,7 @@ ApplyInputGradientsChannelsLastRowMov(
 }

 template <typename T, typename PT, typename opmath_t>
-inline typename std::enable_if<!std::is_same<T, opmath_t>::value, void>::type
+inline std::enable_if_t<!std::is_same_v<T, opmath_t>, void>
 ApplyInputGradientsChannelsLastRowMov(
    const T* dY_data,
    const T* X_data,
--- a/aten/src/ATen/native/cuda/Resize.cpp
+++ b/aten/src/ATen/native/cuda/Resize.cpp
@ -30,7 +30,7 @@ void resize_bytes_cuda(StorageImpl* storage, size_t size_bytes) {
  c10::cuda::CUDAGuard guard(device.index());
  at::DataPtr data = allocator->allocate(size_bytes);
  if (storage->data_ptr()) {
-    at::globalContext().lazyInitCUDA();
+    at::globalContext().lazyInitDevice(c10::DeviceType::CUDA);

    C10_CUDA_CHECK(
        cudaMemcpyAsync(
--- a/aten/src/ATen/native/cuda/layer_norm_kernel.cu
+++ b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
@ -1374,7 +1374,7 @@ std::tuple<Tensor, Tensor, Tensor> layer_norm_cuda(
  for (const auto idx: c10::irange(axis)) {
    stat_shape.push_back(input_shape[idx]);
  }
-  for (const auto C10_UNUSED idx: c10::irange(axis, input.dim())) {
+  for (C10_UNUSED const auto idx: c10::irange(axis, input.dim())) {
    stat_shape.push_back(1);
  }

--- a/aten/src/ATen/native/cudnn/Conv_v8.cpp
+++ b/aten/src/ATen/native/cudnn/Conv_v8.cpp
@ -74,7 +74,7 @@ cudnn_frontend::Tensor getTensorDescriptorWithTypeVirtual(
  // Ubuntu-22+ if `libnvrtc.so` is not found on the system, which strictly
  // speaking is not necessary for usecases below See
  // https://github.com/pytorch/pytorch/issues/97041
-  static C10_UNUSED auto cudnn_cnn_infer_handler = [] {
+  C10_UNUSED static auto cudnn_cnn_infer_handler = [] {
    void* handle = dlopen("libcudnn_cnn_infer.so.8", RTLD_LAZY);
    char* err = dlerror();
    if (!handle) {
--- a/aten/src/ATen/native/mkldnn/Linear.cpp
+++ b/aten/src/ATen/native/mkldnn/Linear.cpp
@ -339,7 +339,7 @@ Tensor mkldnn_linear_pointwise_binary(
 #if AT_MKL_ENABLED()
 #include <mkl.h>

-static Tensor mkl_linear(
+Tensor mkl_linear(
    const Tensor& self,
    const Tensor& mkl_weight_t,
    const Tensor& origin_weight_t,
--- a/aten/src/ATen/native/mkldnn/Linear.h
+++ b/aten/src/ATen/native/mkldnn/Linear.h
@ -22,6 +22,17 @@ C10_API Tensor mkldnn_linear_pointwise_binary(
    const std::optional<Tensor>& bias_opt,
    c10::string_view attr);

+#if AT_MKL_ENABLED()
+
+C10_API Tensor mkl_linear(
+    const Tensor& self,
+    const Tensor& mkl_weight_t,
+    const Tensor& origin_weight_t,
+    const std::optional<Tensor>& bias_opt,
+    const int64_t prepack_batch_size);
+
+#endif// AT_MKL_ENABLED
+
 } // namespace native
 } // namespace at

--- a/aten/src/ATen/native/mps/MPSGraphSequoiaOps.h
+++ b/aten/src/ATen/native/mps/MPSGraphSequoiaOps.h
@ -31,8 +31,15 @@ typedef NS_ENUM(NSInteger, MTLMathMode)
    MTLMathModeFast = 2,
 };

+typedef NS_ENUM(NSInteger, MTLMathFloatingPointFunctions)
+{
+    MTLMathFloatingPointFunctionsFast = 0,
+    MTLMathFloatingPointFunctionsPrecise = 1,
+};
+
@interface MTLCompileOptions()
@property (readwrite, nonatomic) MTLMathMode mathMode;
+@property (readwrite, nonatomic) MTLMathFloatingPointFunctions mathFloatingPointFunctions;
@end

 #endif
--- a/aten/src/ATen/native/mps/OperationUtils.mm
+++ b/aten/src/ATen/native/mps/OperationUtils.mm
@ -541,18 +541,9 @@ Placeholder::Placeholder(MPSGraphTensor* mpsGraphTensor,
    MPSShape* mpsShape = getMPSShape(_tensor);
    MPSShape* mpsStrides = getMPSShape(_tensor.strides());

-    IntArrayRef baseShape;
-    if (src.is_view()) {
-      baseShape = src._base().sizes();
-    } else {
-      baseShape = getIMPSAllocator()->getBufferShape(src.storage().data());
-    }
-    int flattenedShaped = 1;
-    for (const auto i : c10::irange(baseShape.size())) {
-      flattenedShaped *= baseShape[i];
-    }
-    MPSShape* mpsBaseShape = @[ @(flattenedShaped) ];
-    MPSNDArrayDescriptor* srcTensorDesc = [MPSNDArrayDescriptor descriptorWithDataType:dataType shape:mpsBaseShape];
+    auto storage_numel = src.storage().nbytes() / src.element_size();
+    MPSNDArrayDescriptor* srcTensorDesc = [MPSNDArrayDescriptor descriptorWithDataType:dataType
+                                                                                 shape:@[ @(storage_numel) ]];
    srcTensorDesc.preferPackedRows = YES;
    MPSNDArray* srcNDArray = [[[MPSNDArray alloc] initWithBuffer:srcBuf
                                                          offset:src.storage_offset() * src.element_size()
@ -848,7 +839,10 @@ id<MTLLibrary> MetalShaderLibrary::getLibrary(const std::initializer_list<std::s
 }

 id<MTLLibrary> MetalShaderLibrary::compileLibrary(const std::string& src) {
-  static const char* fast_math = std::getenv("PYTORCH_MPS_FAST_MATH");
+  static auto fast_math = []() {
+    auto val = std::getenv("PYTORCH_MPS_FAST_MATH");
+    return val && std::stoi(val) != 0;
+  }();
  NSError* error = nil;
  MTLCompileOptions* options = compile_options;
  if (!options) {
@ -856,7 +850,15 @@ id<MTLLibrary> MetalShaderLibrary::compileLibrary(const std::string& src) {
    // Need 3.0 for atomic oprations, 3.1 introduces bfloat support
    [options setLanguageVersion:is_macos_13_or_newer(MacOSVersion::MACOS_VER_14_0_PLUS) ? MTLLanguageVersion3_1
                                                                                        : MTLLanguageVersion3_0];
-    [options setFastMathEnabled:(!fast_math || std::stoi(fast_math) == 0) ? NO : YES];
+    if (is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS)) {
+      options.mathMode = fast_math ? MTLMathModeFast : MTLMathModeSafe;
+      options.mathFloatingPointFunctions =
+          fast_math ? MTLMathFloatingPointFunctionsFast : MTLMathFloatingPointFunctionsPrecise;
+    } else {
+      C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wdeprecated-declarations")
+      [options setFastMathEnabled:fast_math ? YES : NO];
+      C10_DIAGNOSTIC_POP()
+    }
  }

  const auto str = [NSString stringWithCString:src.c_str() encoding:NSASCIIStringEncoding];
--- a/aten/src/ATen/native/mps/operations/Convolution.mm
+++ b/aten/src/ATen/native/mps/operations/Convolution.mm
@ -167,12 +167,7 @@ static Tensor _mps_convolution_impl(const Tensor& input_t_,

  // TODO: MPS convolution kernel currently does not support output channels > 2^16
  for (auto elem : output_t.sizes()) {
-    TORCH_CHECK_NOT_IMPLEMENTED(
-        elem <= (1 << 16),
-        "Output channels > 65536 not supported at the MPS device. ",
-        "As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` ",
-        "to use the CPU as a fallback for this op. WARNING: this will be slower than running natively ",
-        "on MPS.");
+    TORCH_CHECK_NOT_IMPLEMENTED(elem <= (1 << 16), "Output channels > 65536 not supported at the MPS device. ");
  }

  convolution_shape_check(c, input, weight, output, padding, stride, dilation, groups);
@ -378,12 +373,7 @@ static Tensor mps_convolution_backward_input(IntArrayRef input_size,

  // TODO: MPS convolution kernel currently does not support output channels > 2^16
  for (auto elem : grad_output_t.sizes()) {
-    TORCH_CHECK_NOT_IMPLEMENTED(
-        elem <= (1 << 16),
-        "Output channels > 65536 not supported at the MPS device. ",
-        "As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` ",
-        "to use the CPU as a fallback for this op. WARNING: this will be slower than running natively ",
-        "on MPS.");
+    TORCH_CHECK_NOT_IMPLEMENTED(elem <= (1 << 16), "Output channels > 65536 not supported at the MPS device. ");
  }

  TORCH_CHECK(isFloatingType(grad_output_t.scalar_type()), "Convolution is supported only for Floating types");
--- a/aten/src/ATen/native/mps/operations/Normalization.mm
+++ b/aten/src/ATen/native/mps/operations/Normalization.mm
@ -718,10 +718,15 @@ std::tuple<Tensor, Tensor, Tensor> batch_norm_backward_mps(const Tensor& grad_ou
                                                                  secondaryTensor:epsilonTensor
                                                                             name:nil];
 #ifdef __MAC_15_0
-          rsqrtTensor = [mpsGraph reciprocalSquareRootWithTensor:varianceEpsTensor name:nil];
-#else
-          rsqrtTensor = [mpsGraph reverseSquareRootWithTensor:varianceEpsTensor name:nil];
-#endif
+          if (is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS)) {
+            rsqrtTensor = [mpsGraph reciprocalSquareRootWithTensor:varianceEpsTensor name:nil];
+          } else
+#endif // __MAC_15_0
+          {
+            C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wdeprecated-declarations")
+            rsqrtTensor = [mpsGraph reverseSquareRootWithTensor:varianceEpsTensor name:nil];
+            C10_DIAGNOSTIC_POP()
+          }
          MPSGraphTensor* bnForwardTensor = [mpsGraph multiplicationWithPrimaryTensor:xMinusMean
                                                                      secondaryTensor:rsqrtTensor
                                                                                 name:nil];
@ -747,10 +752,15 @@ std::tuple<Tensor, Tensor, Tensor> batch_norm_backward_mps(const Tensor& grad_ou
                                                                    secondaryTensor:epsilonTensor
                                                                               name:nil];
 #ifdef __MAC_15_0
-            rsqrtTensor = [mpsGraph reciprocalSquareRootWithTensor:varianceEpsTensor name:nil];
-#else
-            rsqrtTensor = [mpsGraph reverseSquareRootWithTensor:varianceEpsTensor name:nil];
-#endif
+            if (is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS)) {
+              rsqrtTensor = [mpsGraph reciprocalSquareRootWithTensor:varianceEpsTensor name:nil];
+            } else
+#endif // __MAC_15_0
+            {
+              C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wdeprecated-declarations")
+              rsqrtTensor = [mpsGraph reverseSquareRootWithTensor:varianceEpsTensor name:nil];
+              C10_DIAGNOSTIC_POP()
+            }
          }

          gradInputTensor = [mpsGraph multiplicationWithPrimaryTensor:unitTensor secondaryTensor:rsqrtTensor name:nil];
--- a/Show More
+++ b/Show More